A State of Art Review on Methodologies of Occupancy Estimating in Buildings from 2011 to 2021

Abstract

Occupancy information is important to building facility managers in terms of building energy efficiency, indoor environmental quality, comfort conditions, and safety management of buildings. When combing the distribution characteristics of the literature, it is found that the field of estimating occupancy counts is a very active area. Researchers from various countries have undertaken extensive explorations to obtain more research results. In this survey, the commonly used occupancy measurement systems and algorithms are described. Through the analysis and research of different occupancy measurement systems and algorithms, their advantages, disadvantages, and limitations are summarized, so that researchers can use them selectively. As for how to choose the method of estimating occupancy counts, suggestions are given in terms of the range of people, accuracy, cost, and privacy. There are still many pressing issues relating to high-density crowd occupancy counting, complex environmental impact, and system robustness. According to the current research progress and technology development trend, the possible future research directions are pointed out. The innovation of this review is the quantitative analysis of the selection of occupancy measurement systems for different ranges of people, and the occupancy counting accuracy situation of different measurement systems and algorithms. It provides more informed opinions on the selection of practical applications. It can be used by other researchers as a starting point for their research and/or project work.

1. Introduction

In recent years, the trend of urban building intelligence has been increasing. More effective control and management of buildings can be implemented based on information about the number of people in a certain room, helping to achieve the attributes of safety, efficiency, convenience, energy saving, environmental protection, and health of the building. Determining the number of people in a given space has played an extremely important role in many fields and has been widely used in many places. People-counting is an important link in intelligent video surveillance systems. It has a wide range of applications and commercial value in banks, railway stations, shopping centers, schools, and other places [1]. Real-time estimation of occupancy rates in the housing of the elderly or the disabled can be used to improve the safety of residents [2]. In places where the density of people needs to be controlled, the number of people counted by an identification technology can provide a reference for the regulatory department. In addition, information on the number of people in the room plays a very important role in achieving energy efficiency in buildings. Occupancy counting has a great impact on IEQ and the energy consumption of buildings. At the same time, the existence of humans will also put forward demands on the environment, such as the need for more suitable temperature, humidity, light, air, etc. [3,4]. Operators can configure the HVAC, lighting, and electrical appliances to turn on according to the number of occupants. During the idle period, the system can be set to the minimum flow setting or completely closed. Nowadays, more and more studies have begun to pay attention to occupancy counting. A full grasp of information about occupancy counting can bring substantial conveniences to life.

Among the studies of occupancy information, some studies focus more on whether the rooms are occupied without estimating the exact number of people [5,6,7,8]. This type of building occupancy information is called occupancy detection. There are also some studies dedicated to people detection and tracking [9,10,11,12,13]. This type of occupancy information is called occupancy tracking. More accurate occupancy counting will play a greater role in saving building energy consumption. It can not only obtain information about presence or absence, but also focus on the number of occupants in a particular room or zone [14]. There have been many studies on occupancy counting. Commonly used occupancy measurement systems include environmental sensors, infrared, wireless communication equipment, images/video-based methods, etc. [15]. To estimate the number of people more accurately and effectively, some studies have combined the advantages and disadvantages of various sensors to detect occupancy through sensor fusion. Most of the current occupancy counting methods can be split into vision-based techniques and data-based methods [16]. These two methods are applied to two different occasions for use according to their characteristics. One is the need to directly detect each person in the scene and then count the people. The other is the need to establish the relationship between some characteristic parameters and the number of people in order to estimate the number of people. Counting people in a crowded surveillance environment is a challenging task, and the analyst is usually more inclined to establish relationships to estimate the number of people, because it is difficult to effectively segment each person in a highly crowded scene.

Hobson et al. [17] applied a variety of different sensors/sensor fusions for occupancy counting. Multiple linear regression and ANN models were developed to estimate occupancy counts by using Wi-Fi device counts, CO₂ concentrations, PIR motion detector triggers, plug load, and lighting load data from an academic office building. By comparing the estimates with ground truth data, it can be seen that different sensors, sensor combinations, and model formalisms for occupancy-count estimation perform differently when estimating occupancy or counting. It is shown that the accuracy of the estimated occupancy or counting has a large correlation with both the sensors used and the algorithmic model. At present, there are already comprehensive review articles on occupant sensing in the built environment. Some of them focus mainly on the overview of occupancy measurement systems [14,18]. The occupant-sensing technology in occupancy detection is discussed in terms of sensors [19]. Choi et al. [20] performed a comprehensive and structural literature review of vision-based occupant information systems. Kouyoumdjieva et al. [21] reviewed non-image-based approaches for counting people. There are also studies focusing on the classification and discussion of mathematical models of occupancy detection algorithms [22,23]. Although some papers have also synthesized occupancy measurement systems and algorithms for the review [24,25,26], no review has quantified the analysis in terms of experimental accuracy and whether the conclusions lack support in the data. There is also no review addressing the preference of different population ranges in choosing occupancy measurement systems. In view of the limitations of the above studies, this paper presents a systematic overview of the occupancy measurement systems and algorithms used for estimating occupancy counts. It will also provide more informative assessments of the selection of occupancy systems and algorithms for practical applications, especially for different numbers of measurement people and the accuracy of use. This paper reviews the studies on occupancy counting in recent years. The contributions are as follows:

(1) We focus on estimating occupancy counts in buildings, and review many recent studies, concentrating on studies within the last decade.

(2) The methods used in the systematic literature review and the distribution characteristics of the literature are sorted out.

(3) The commonly used occupancy measurement systems and occupancy counting algorithms, as well as the aspects that need to be considered when the question is put into application, are analyzed and discussed.

(4) Quantitative analysis to provide guidance on the selection of occupancy measurement systems and algorithms in terms of accuracy, number of personnel, cost, privacy, etc. is conducted.

(5) The general problems faced by current studies are put forward, and the review looks forward to future studies. The paper is organized as follows: Section 1 introduces the background and purpose of this review. Section 2 describes the review methodology. Section 3 focuses on the distribution characteristics of literature, occupancy measurement systems and occupancy counting algorithms. Section 4 discusses how to choose the method of estimating occupancy counts and presents the challenges faced by current research. Finally, Section 5 concludes this review work and proposes future development directions. A schematic overview of this paper can be found in Figure 1.

2. Methodology

Based on an array of pre-collected reviews and dissertations in the field of occupancy information, we focused on reading the English reviews, understanding the frontiers, difficulties, and innovations, and collecting keywords. The first step in creating the systematic literature review is to define the scope of the study. As for the selection of research literature time scope, through literature collection, we found that the number of papers in the relevant literature was relatively small before 2011, when the development of research was relatively slow, and the technology update iteration was not fast. After comprehensive consideration, to ensure the integrity of the article collection period and the timeliness of the research, the literature from 2011 to 2021 was selected as the relevant time period. The purpose of this review is to sort out and compare the studies on occupancy counting over the past ten years. This review includes studies published in the following databases from 2011 to 2021: Web of Science, Scopus and Engineering Village.

After defining the scope, we proceeded to collect papers relevant to the systematic literature review. The first and foremost task is to determine the search keywords. The keywords used in the literature search were: “occupancy counting”, “occupancy detection”, “people counting”, “occupancy measurement”, “occupancy rate”, “occupancy estimation”, and “occupancy sensing”. Different keywords were connected by “AND” or “OR” to indicate the relationship between keywords. To integrate the literature, we searched in different databases and removed duplicates. In the initial search, a total of 769 highly relevant publications were selected based on the logical combination of keywords. The titles of the first set of selected publications were screened according to their research background. Publications that are less relevant to occupancy research were excluded. Of those, 150 publications remained, and their abstracts were appropriately reviewed. The review of the abstract included: (i) excluding theoretical research publications that only carry out sensor and algorithm improvement; (ii) excluding publications that have nothing to do with building occupancy information research; (iii) excluding publications that only summarize human detection methods without experimental research. After exclusion according to these criteria, 88 publications remained. When reading the selected 88 publications, it was found that some publications only focused on whether the building was occupied, and did not mention the identification of the specific number of people. These publications were eliminated, leaving 78 publications. These 78 publications were used as the core literature of this review and will be reviewed in detail later. The entire publication review flowchart with the number of studies identified and included/excluded is shown in Figure 2.

3. Review Results

After going through the process of paper collection, all the papers were then analyzed thoroughly for the purpose of summarization. Important facts from each paper were taken and used in our systematic literature review. This paper reviews 78 core pieces of literature, including the occupancy measurement systems and methods used in estimating occupancy counts, as well as estimation accuracy obtained through experiments, and so on.

Table 1. The review results in detail in the chronological order of the literature publication.

Study	Years	Location	Sensors	Methods	Accuracy	People Number	Remarks
[27]	2021	Poland	Camera	CNN	94.1%	5
[28]	2021	Korea	Camera	YOLO	NRMSE H: 0.0918 L: 0.0435	13	major over/under counting cases
[29]	2021	UK	Infrared	Boosting algorithm Regression	H: 98.43% L: 93.75%	4	thermal sensor
[30]	2021	Sweden	Infrared	Multi-Bernoulli Target Tracking Approach	H: 98% L: 86%	-	IR cameras
[31]	2021	India Singapore Switzerland	Wireless Communication		87%	97
[32]	2021	USA	Time-Of-Flight Sensors	Zone Counting	H: 100% L: 91.7%	3
[33]	2021	China	Infrared	Gaussian Distribution Model	85%	65	enter or leave in turn
[34]	2021	Spain	Wireless Communication	Footprint Calculation Algorithm	95%	700
[35]	2021	USA	Environment Sensor	Physics-Informed Pattern-Recognition Machine	97%	2
[36]	2021		Infrared	RLS	89%	4	thermopile infrared sensing
[37]	2021	Denmark	Infrared	Binarization	87.5%	19	passive-infrared
[38]	2021	USA	Wireless Communication	Probability Distribution Function	96.3%	27	stationary crowd
[39]	2021	USA	Infrared	Gaussian Naive Bayes KNN SVM Random Forest	99%	8	thermal imaging sensors
[40]	2020	USA	Wireless Communication	Iterative Dynamic Programming Algorithm Trace Concatenating Algorithm	86%	4	static crowd counting
[41]	2020	USA	Camera	YOLO	H: 83.95% L: 57.52%	25
[42]	2020	China	Wireless Communication	SVM	H: 86.4% L: 78.2%	-	without upper bound of people counted
[43]	2020	Germany	Radar Sensor	Adaptive OS-CFAR Peak Detection Algorithm Vital Sign Verification Algorithm	85.4%	4	60 GHz mm-wave radar
[44]	2020	USA	Wireless Communication	CNN	82%	5	RFID
[45]	2020	China	Wireless Communication	CNN	94.5%	6	visitor counting system
[46]	2020	UK	Wireless Communication	CNN	98.14%	4	low-cost passive Wi-Fi radar
[47]	2020	USA	Infrared	ANN	91%	3	PIR sensors
[48]	2020	China	Infrared	HMM Softmax Regression Model	H: 99.44% L: 80%	7
[49]	2020	China Canada	Environment Sensor	Multi-Grained Scanning Cascade Forests	MAE: 0.153	4
[50]	2020	UK	Infrared	Adaboost	98.2%	4	infrared thermal sensor array
[51]	2020	USA	Infrared	Baseline Algorithm Multi-Person Algorithm	MAE H: 1.431 L: 0.003	133	low-resolution thermal sensor
[52]	2020	Mexico UK	Environment Sensor	Decision Tree KNN SVM	97%	4
[53]	2020	China	Camera	CNN	70%	80	estimate the real-time load complex scenes a small size and a fast model inference speed
[54]	2019	USA	Wireless Communication	Random Forest ANN LSTM	RMSE H: 4.62 L: 1.20	74	protect privacy.
[55]	2019	Denmark	Wireless Communication	Multi-Modal Fusion Algorithm	RMSE H: 1.39 L: 0.87	600
[56]	2019	Belgium	Camera	Deep Learning	H: 96.7% L: 42.8%	4	privacy preserving
[57]	2019	USA	Wireless Communication		97%	-	the first intelligent IoT sensor platform
[58]	2019	UK	Infrared	Human Detection and Direction Algorithm			occupancy monitoring unit thermal imaging
[59]	2019	Denmark Norway	Environment Sensor	Stochastic Differential Equations	77%	5	influence of window openings has not been investigated
[60]	2019	Italy	Wireless Communication	Logistic Regression Classifier	H: 97% L: 53%	132	low cost different levels of crowding
[17]	2019	Canada	Environment Sensor Wireless Communication Infrared	Linear Regression ANN	H: 97.5% L: <0.1%	72	sensor fusion
[61]	2019	USA	Infrared	Window Size Algorithm Connected Component Algorithm	H: 100% L: 17%	3	low-resolution thermal sensors without causing any discomfort or privacy issues
[62]	2019	USA	Environment Sensor	Inverse Modeling Algorithm	-	-	develops a novel inverse modeling method
[63]	2019	Korea	Environment Sensor	Decision Tree SVM ANN	H: 97.19% L: 90.52%	-	reselection of input variables each season and situation
[64]	2018	China	Environment Sensor	Parameter Model Kalman Filtering	H: 88.81% L: 71.22%	4
[65]	2018	Australia	Environment Sensor	Semi-supervised Domain Adaptation Method	H: 75.25% L: 55.74%	230	minimal training data
[16]	2018	Singapore Australia	Environment Sensor Camera	SVM Physical and Statistical Model	H: 97.2% L: 26.4%	200	four different occupancy estimation methods
[66]	2018	Australia	Wireless Communication	Logistic Regression SVM LDA	H: 84% L: 79%	250	without incurring cost and effort
[67]	2018	China USA	Wireless Communication	Markov Based Feedback Recurrent Neural Network Algorithm	H: 93.9%. L: 80.9%	19	more reliable and sensitive
[68]	2018	USA	Camera	Cell-tracking Methodology	98.12%	8	the CellProfiler image processing software
[69]	2018	USA Singapore	Wireless Communication	Transfer Kernel Learning Based Classifier	92.8%	11	IoT preserve the privacy robust to temporal and environmental
[70]	2018	Korea	Environment Sensor	MCMC	RMSE H: 3.91 person L: 0.11 person	25	inter-connected rooms multi-room office building
[71]	2018	China USA	Environment Sensor Wireless Communication	KNN SVM ANN	MAE H: 3.0 L: 1.7	25	data fusion
[72]	2017	USA	Device	Linear Regression	50–80%	350	electricity consumption
[73]	2017	Canada	Device	GEP	>90%	-	data collected by PC single-person office occupancy
[74]	2017	Singapore	Camera	SVM	92%	220	people-occupancy estimation system
[75]	2017	China	Camera Environment Sensor	Conservation Equation	-	-	CO2 data revise the results of video detection
[76]	2017	Singapore	Environment Sensor	Dynamic Physical Models SVM ANN	RMSE H: 12.8 L: 12.1	200	prediction of occupancy counts in a large space
[77]	2017	Singapore	Environment Sensor	ELM	>96%	14	high accuracy faster
[78]	2017	Poland	Environment Sensor	KNN LDA	ME H: 0.79 L: 0.45	42	determine the duration of occupancy periods
[79]	2017	Belgium	Environment Sensor	HMM	H: 92.54% L: 19.1%.	-
[80]	2017	Australia	Infrared	K* C4.5 MLP	H: 98.15% L: 83.68%	3	K* is an adaption of KNN thermal images low-resolution thermal sensors
[81]	2017	China	Camera	SVM CNN K-means	95.3%	12	low computational cost high precision
[82]	2017	China USA Singapore	Wireless Communication	ELM	98.85%	32	accurate, reliable, cost-effective and non-intrusive
[2]	2016	Australia	Infrared	K*(KNN) C4.5 ANN SVM Linear Regression Naive Bayes	H: 82.56% L: 49.74%	3	thermal detector array privacy-preserving low-cost non-invasive energy efficient
[83]	2016	Korea	Environment Sensor	Decision Tree HMM	H: 93.2% L: 85.0%	5	occupancy prediction
[84]	2016	Singapore	Environment Sensor	ELM	94%	28	locally smoothed data
[85]	2016	Denmark	Camera	Waterfilling Algorithm	H: 100% L: 98.4%	3
[86]	2016	UK Singapore China	Wireless Communication	CRF SVM	ME: H: 1.615 L: 0.560	25	inference accuracy and robustness
[1]	2016	China	Camera	Adaboost Algorithm CNN SVM	H: 97% L: 50%	82	low resolution body occlusion unconstrained imaging viewpoints
[87]	2016	UK USA	Infrared	HMM	H: 99% L: 59%	14	low-cost
[88]	2016	Singapore	Environment Sensor	ELM SVM ANN KNN LDA CART	H: 0.7390 L: 0.6100	10
[89]	2016	Singapore	Camera	SVM	99%	-	public locations
[90]	2015	Italy	Depth Sensor	Adaboost Classifier	MRE H: 6.4% L: 1.4%	40	higher level of privacy preservation
[91]	2015	Netherlands	Device		100%	13	chair sensors viable, low cost, reliable
[92]	2014	Finland	Device Camera		H: 97% L: 80%	150
[93]	2014	China	Camera	SVM		8	PTZ camera detection and tracking robust for counting people
[94]	2014	Italy Switzerland	Camera	CNN	RMSE H: 8.55 L: 6.46	80
[95]	2013	China	Camera	Haar–Adaboost Classifiers SVM	H: 92.03% L: 86.21%	8	fusion of vision sensors
[96]	2013	USA	Infrared	KNN ANN Linear Regression	RMSE 0.35 persons	3	low-cost low-power thermal sensor
[97]	2013	Denmark	Infrared	Automatic Threshold Method Based on Maximum Entropy	95.56%	13	thermal camera
[98]	2012	Denmark	Infrared	Automatic Threshold Method Based on Maximum Entropy	H: 99.83% L: 88.24%	15	thermal camera prevent privacy
[99]	2012	Eindhoven	Infrared	Direction-based Algorithm Probabilistic Distance-based Algorithm		3	inexpensive and easily installed
[100]	2011	France	Camera	Classifiers	H: 93% L: 83%	2	detect human

shows the review results in detail in the chronological order of the literature publication.

3.1. Distribution Characteristics

The distribution characteristics of literature will be discussed from two dimensions; one is the publication year distribution of the literature, and the other is the geographic distribution of the literature. The overall development of estimation of occupancy counts research can be concluded based on the discussion of the distribution characteristics of the research literature. For example: Whether estimating occupancy counts becomes popular in these years? Whether the system attracts the attention of both developed and developing countries?

3.1.1. Distribution by Year of Publication

The distribution of the literature by year of publication gives a more objective indication of how old or new these literature studies are. It is generally more effective to determine the range of years of the literature to be studied based on the speed of the corresponding technological iteration. Studies in recent years will show the current status of research and future trends. According to the literature collection, the research literature before 2011 is relatively scattered and small in number. Moreover, to ensure the timeliness of this piece of review, this review did not trace the research literature too far back. In this paper, the studies in the recent ten years are selected for analysis and discussion. Figure 2 shows the publication year distribution of the literature of this review. Although the year range of this study was 2011–2021, most of the selected literature, i.e., 67 of the 78 pieces of literature, were published in 2016–2021. Figure 3 clearly shows the unprecedented increase in the number of studies in the field of estimating occupancy counts since 2016. It reflects that with the popularization and development of intelligent buildings, both academia and industry pay more and more attention to the estimation of indoor populations. Research scholars and institutions in various countries have undertaken a significant amount of exploration and efforts in these aspects and obtained more research results.

3.1.2. Distribution by Authors’ Nationality

The reviewed studies were globally distributed and did not concentrate on a specific region. The distribution of authors in the remaining studies is discussed, excluding some studies whose authors’ country information cannot be inferred from literature. Figure 4 shows the geographical distribution of the literature. After sorting and analyzing, it can be seen that all the literature comes from 21 countries. These countries are concentrated in four continents: Asia, Europe, Oceania, and North America. By analyzing the geographical distribution of literature, we can understand which regions are interested in estimation of occupancy count research. After analysis, it is found that a vast number of scientific researchers and related investigators in the world have a strong interest in estimating occupancy counts, because it has a wide range of application prospects and potential economic value. Especially in China, the USA and other countries, a large number of research projects have been carried out. In the ranking according to the number of papers reviewed in this review, the research in China and the USA ranks at the forefront. There are as many as 37 pieces of research literature in China and USA, accounting for about 47% of the total number of articles. This shows the general interest of developing countries (headed by China) and developed countries (headed by the USA) in the study of estimating occupancy counts.

3.2. Review of Occupancy Measurement Systems

This section describes the most widely used data collection techniques in recent occupancy estimation studies. We will give a comprehensive overview of occupancy measurement systems used in the literature. It will provide a reference for the use of sensors to estimate occupancy counts. After statistical analysis, the commonly used occupancy measurement systems are divided into four categories: camera, environmental sensors, wireless communication, and infrared. The remaining less-frequently used sensors are automatically classified into another category. The literature referred to in this paper most frequently looks to methodologies using cameras and environmental sensors, both accounting for 24% of the total studies. The use of wireless communication technologies accounted for 23%, followed closely by the use of infrared technologies at 21%. In addition to these four categories, the remaining studies account for a total of 8%. Figure 5 shows the usage proportion of several occupancy measurement systems. Next, the occupancy estimation and detection literature will be classified according to the sensors used. At the same time, the advantages and limitations of various occupancy measurement systems will be compared.

3.2.1. Camera

Using a camera to estimate occupancy counts is generally based on methods incorporating images or videos. Because of its high accuracy, these methods are often used for building occupancy estimation and detection. The common method of using the camera for occupancy detection is to detect and track the heads, faces, body contours, or motions of occupants. By analyzing the images taken by the camera and extracting human body features, accurate occupancy information can be obtained. However, image/video detection sometimes encounters some difficulties. For example, there are generally many obstacles in the monitoring environment, the body is only partially exposed in the video, and the detection method using a single human feature often fails to detect the human body in the actual scene. Not only that, but the use of a camera usually requires good lighting conditions. In addition, the invasion of the privacy of the occupants must be another key issue in the use of cameras. However, few studies have focused on the use of a more recent type of camera, ToF cameras, which have good privacy protection [68]. This is likely due to the fact that, in terms of cost, operators prefer to use existing surveillance video and try to avoid installing new cameras. In a study by Meng et al. [53], a camera was used to collect image information inside a building, and computer vision and deep learning detection technology were applied. The study was based on CNN establishing an end-to-end building-space staff-load dynamic-estimation model to realize real-time detection of the number of indoor personnel and estimate personnel changes. Chandran et al. [74] proposed a single PTZ Camera Based People-Occupancy Estimation System (PCBPOES). In this method, the PTZ camera conducted effective surveillance by dividing the area into multiple areas, and detecting people’s heads to estimate the number of people in the area. Gade et al. [98] presented a system for automatic analysis of the occupancy of sports areas. The system used a thermal imaging camera to take images, which can be quite precise in distinguishing between an empty area, a few people, or many people. This method of using thermal cameras to automatically detect people could not only find the number of people and their positions on the court, but the system also did a good job of protecting privacy. Using camera-based methods has good results in terms of estimation accuracy, but, due to additional hardware costs and privacy issues, the use of these methods is limited [59]. According to recent studies, the realization of crowd counting based on cameras still faces some challenges. Due to some limitations, such as the impact of the camera’s viewing angle, the system cannot meet certain requirements. There are also angular distortions that can lead to less-accurate results. In a crowded environment, the forced use of the system is difficult, and it will inevitably obscure people and distort the viewing angle. In terms of system scalability, distortion is caused due to the lack of additional training in different spaces or locations. Most of the models used lack training capacity, and as a result, are only familiar with a specific scenario, so the system cannot be implemented in different scenarios. Due to a lack of data training, when the system is to be implemented in a crowded environment, accurate data cannot be obtained. This causes the final crowd count estimate to be underestimated or overestimated in the crowd image.

3.2.2. Environmental Sensors

Environmental sensors mainly include CO₂, temperature, humidity, humidity ratio, light, etc. Among all environmental sensors, CO₂ has been proven to be the most effective sensor for occupancy estimation and detection [65]. Some studies use only CO₂ data to estimate building occupancy, while others use multiple environmental sensor fusion methods. Occupancy counting through the CO₂ sensor is a very promising method, because indoor CO₂ concentration is a sign of human existence [76]. In addition, the CO₂ sensors are cheap, small, non-invasive, and nonterminal-based. Franco et al. [101] conducted experiments in different classrooms on a university campus through different periods of the year and under different types of activities. The study proved the correlation between the measured value of CO₂ and the number of occupants.

Generally speaking, the estimation of occupancy counts based on the CO₂ mass balance equation is the most typical [59,64]. Studies using machine algorithms are also very common [63,77,83]. Li et al. [62] developed a set of new inverse modeling algorithms that can solve highly uncertain and difficult-to-measure building parameters. According to model assumptions and using the measured zone air temperature, humidity and/or CO₂ concentration, the algorithms are used to solve different unknown parameters, such as the number of people. Rahman et al. [70] developed a method for identifying the number and distribution of people in a multi-room office building based on CO₂ concentration. It used Bayesian inference to estimate the occupancy distribution of multi-room office buildings. A Markov chain Monte Carlo algorithm was used to estimate the occupancy rate of each room in real time, based on the indoor CO₂ concentration.

There are still some issues to consider in the future using environmental sensor measurements to estimate occupancy counts. For example, since environmental characteristics are very sensitive to changes in the environment itself, a sensing system using environmental sensors requires higher accuracy. Human behavior will also have a great impact on the accuracy of the estimation. For the most commonly used CO₂ sensors, the following considerations need to be made when estimating occupancy counts. According to current level of activities, diets and body sizes, human CO₂ production rates vary greatly from person to person. [102]. Depending on the ventilation conditions and the number of activities performed at each location, the CO₂ concentration level will vary from one location to another [103]. The scheme based on CO₂ has low accuracy in estimating a large number of people, and is only suitable for roughly estimating the number of people. However, the dynamic response of this method is slow, because it takes some time for the CO₂ concentration to change if the occupant enters or leaves the room. If the people entering the room leave the room so quickly that the CO₂ concentration in the room hardly changes, their number may not be considered in the occupancy estimation. In addition, according to the ventilation, the CO₂ concentration varies from room to room, and a general model to deduce the occupancy rate with the measured CO₂ level may not be found.

3.2.3. Wireless Communication

The use of wireless communication technology to estimate occupancy counts is mainly through Wi-Fi, Bluetooth, BLE, RFID, and so on [44]. In recent years, with the continuous improvement of Wi-Fi infrastructure applications and ubiquitous Wi-Fi-enabled mobile devices, it has become possible to use COTS Wi-Fi routers and mobile devices carried by passengers for occupancy detection. At the same time, the number of Bluetooth-enabled personal devices (smartphones, smartwatches, wireless headsets, etc.) that people carry has increased significantly. Such devices work according to two main standards: Bluetooth Classic and Bluetooth Low Energy. Compared with Bluetooth Classic technology and Wi-Fi, BLE is quite energy-saving. Many researchers also intend to use BLE to estimate the number of smartphones to understand the number of indoor residents. Longo et al. [60] proposed a system for estimating the occupancy of spaces based on the capture of Wi-Fi probe requests and Bluetooth/BLE management frame. The performance of the system when using Bluetooth, Wi-Fi, or some combination of the two are compared. The indoor and outdoor experimental results prove the effectiveness of the method. Lu et al. [86] proposed a robust occupancy inference system based on commercial Wi-Fi hardware, which can obtain accurate occupancy information based on the existing Wi-Fi infrastructure with minimal user workload. Zou et al. [69] proposed WiFree, Wi-Fi-based device-free occupancy detection and crowd counting method using only COTS Wi-Fi-enabled IoT devices. It can achieve 92.8% people counting accuracy without using the device, while also protecting the privacy of occupants.

Due to the wide application of Wi-Fi signals in indoor environments, using Wi-Fi for occupancy estimation and detection is a good method. But the limitations of Wi-Fi-based system detection are obvious. These systems will encounter some problems when occupancy counting is obtained indirectly by detecting the number of smartphones. For example, passengers may have multiple smartphones; their smartphones may not have Wi-Fi turned on; occupants may go out but forget to bring their phones, etc. In addition to the problems mentioned above, the estimated occupancy count based on BLE requires additional cost and maintenance to deploy BLE facilities in indoor environments.

3.2.4. Infrared

Infrared, sometimes called infrared light, is electromagnetic radiation with wavelengths longer than those of visible light. An infrared-based thermal sensor array is a commonly used method for estimating occupancy numbers. The thermal sensor array offers privacy-preserving, low-cost, and non-invasive features. Naser et al. [50] proposed a non-contact scheme for occupancy estimation using an infrared thermal sensor array, which has the advantages of being low-cost and low-power, and possessing high-performance capabilities. The proposed scheme offers an accurate human heat segmentation technique that extracts human body temperature from a noisy environment. Tyndall et al. [2] used training data obtained from a four by sixteen thermal detector array and a PIR sensor. Machine learning classifiers are used to interpret the raw data from the detector array in order to deduce the number of occupants in the sensor’s field of view. Infrared-based sensors have better advantages in terms of cheapness, simplicity, non-intrusiveness and privacy protection, but they have a limited ability to estimate occupancy counts and are better suited for situations where a small number of people need to be monitored. The PIR sensor is one of the most commonly used occupancy-detection devices. PIR sensors can detect changes in infrared radiation and reflect information derived from the movement of objects. PIR sensors have the advantages of being low cost and running with low energy consumption, while maintaining high reliability, and an easy application to the environment. However, they cannot detect relatively stationary passengers, and the detection range of each sensor is limited. Raykov et al. [87] showed how to use a single PIR sensor combined with machine learning models to solve the counting problem in a room. The system has been successfully tested on data from more than 50 real office meetings with a maximum of 14 people. However, the use of a single sensor has the following problem: Since most infrared radiation is reflected from the human body, people in the monitored environment are easily blocked by others from the field of vision of the single sensor. Wahl et al. [99] used distributed strategically placed PIR sensors and algorithms to estimate the number of people in each office space. It is found that using pairs of inexpensive PIR sensor nodes to estimate the occupancy count not only saves costs, but also, when using distributed sensor information, can partially compensate for errors introduced by PIR masking.

3.2.5. Others

In addition to some of the sensors mentioned above, many other sensors are also used for occupancy estimation and detection. Diraco et al. [90] used a depth sensor for occupancy detection. It can not only reliably count and locate people in a crowded environment, but also, because the depth information does not reveal the identity of the detected person, it can more effectively protect privacy. Labeodan et al. [91] introduced chair sensors in an office building conference room and evaluated their performance. Experiments show that chair sensors have achieved good results in detection accuracy. Each sensor has its advantages and limitations in occupancy estimation and detection. The fusion of multiple types of sensors can improve the performance of occupancy estimation and detection by compensating for the limitations of each sensor. Wang et al. [75] proposed a method to detect room occupant numbers by data fusion of the video and CO₂ concentration: Use camera detection or CO₂ concentration estimation in different time periods. The number of people estimated at night using the CO₂ concentration is used as the initial number of people for video detection. In this way, the cumulative error can be eliminated, and the accuracy of the occupancy rate prediction can be improved. Wang et al. [71] proposed that the fusion of Wi-Fi connections and environmental parameters can further improve the accuracy of occupancy predictions. By applying multiple sensors, the parameters obtained can show the correlation with the number of people, which may be used to improve accuracy. However, there has not been a comprehensive study of which sensor combination can improve accuracy, thus justifying the increase in the cost from installing additional sensing infrastructure.

3.3. Review of Estimating Occupancy Counting Algorithms

Most occupancy estimates use data from one or more sensors. They need to extract features from the data so that subsequent occupancy estimations can be carried out with sufficient accuracy. This section will review the occupancy estimation algorithms used in the literature. Through sorting out, it is found that machine learning is the most commonly used method of estimating occupancy counts. Nearly 76% of these selected papers use machine learning methods to estimate occupancy counts. Some use the form of establishing functional equations to estimate occupancy counts, a methodology that accounts for 8% of the papers reviewed; most are used when using environmental sensors. In addition, other algorithms used account for about 16%. Figure 6 shows the proportion of algorithm usage. In recent years, the application of machine learning in estimating occupancy counts has increased substantially. At the same time, machine learning is a common research hotspot in the field of artificial intelligence and pattern recognition. Its theories and methods are also widely used to solve complex problems in engineering applications and scientific fields. Several popular machine learning methods in this review will be discussed in detail below.

In the field of occupancy estimation, common machine learning models include logistic regression, ANN, Markov Chain Model, Decision Trees, KNN, SVM, etc. Most occupancy studies using machine learning algorithms are based on supervised learning, which tries to learn a function with which to map an input to an output based on example input-output pairs [104]. Deep learning classifiers have been recently utilized in the context of image-based people counting. The concept of deep learning comes from research relating to ANN. A multilayer perceptron with multiple hidden layers is a deep learning structure. Deep learning combines low-level features to form more abstract high-level representations of attributes, categories, or features, in order to find the distributed feature representation of data. Deep learning is a new research direction in the field of machine learning. Deep learning seeks to learn the internal laws and representation levels within sample data. The information obtained in the learning process is of great help to the interpretation of data such as text, images and sounds. What is worth discussing is the famous ability of deep learning models, which can not only automatically learn the pattern between features and the category to be predicted, but also automatically learn the features themselves. We now focus on the literature using machine learning algorithms, and examine statistics on the frequency of specific machine learning algorithms. The five most frequently used machine learning algorithms in the literature were identified; they are SVM, ANN, CNN, KNN, and Regression.

3.3.1. SVM

SVM is a kind of generalized linear classifier that classifies data according to supervised learning. Its decision boundary is the maximum margin hyperplane for learning samples. This algorithm can be used for classification and regression. It works by performing a projection of the data points to a higher dimension, where they can be separated using a line or a hyperplane. The points that lie exactly at the limit of the decision surface are called support vectors [39]. Through minimizing hinge losses, SVM ignores training data that are very close to the actual results but intends to model a boundary that includes as many samples as possible, in order to improve the model’s reliability. In the literature reviewed here, SVM is the most frequently used machine learning algorithm. Zhang et al. [42] introduced the SVM algorithm to seek a decision function identifying the number of people. Finally, the number of people in the crowd flow can be obtained without specific environment calibration. Chandran et al. [74] uses a cascade detector of SVM to detect human heads. In contrast to some machine learning methods, the SVM method works well with high-dimensional data and it is very fast in prediction since the model is affected by only a small number of support vectors [81]. Shih et al. [93] uses a new SVM-based observational measurement, which provides a robust day-and-night occupant tracking and counting performance.

3.3.2. ANN

ANN refers to a complex network structure formed by interconnecting a large number of processing units, i.e., neurons. It is a certain abstraction, simplification, and simulation of the human brain tissue structure and operating mechanism [54]. ANN simulates neurons’ activity with mathematical models. It is an information processing system based on imitating the structure and function of the brain’s neural network. There are multiple layer and single layer versions of ANN. Each layer contains several neurons, and each neuron is connected by a directed arc with variable weights. The network is trained by repeated learning of known information and gradually adjusts to change the neuron connection weights. Eventually, it achieves the purpose of processing information and simulating the relationship between inputs and outputs. Each layer can adjust the level of accuracy by changing the number of neurons [63]. It does not need to know the exact relationship between input and output, and does not need a large number of parameters. It only needs to know the non-constant factor that causes the output change, that is, the non-constant parameter. Therefore, compared with traditional data processing methods, neural network technology has obvious advantages in processing fuzzy data, random data, and nonlinear data. It is especially suitable for systems with large scale, complex structures and unclear information. Andrews et al. [47] accurately detected the presence of people with the combined technology of an ANN. The ANN models provide the ability to learn from the data and thereby make accurate predictions.

3.3.3. CNN

CNN is a typical deep learning model. With the gradual rise of artificial intelligence, deep learning algorithms have gradually attracted attention due to their excellent results in many fields. CNN is one of the most representative algorithms in the field of deep learning. Because it can pay attention to the subtle features of the image and has a very high advantage in image processing, many people-recognition methods based on CNN have emerged [45]. Bao et al. [105] propose a novel people-counting algorithm exploiting CNN using a low radiation impulse radio ultra-wide bandwidth (IR-UWB) radar. They conclude that this algorithm is more stable when applied to situations with obstruction and superposition in scenes with a wider detecting angle and larger detecting range. By training with CNNs, Tang et al. [46] obtained general models for occupancy sensing which provided good estimation accuracy. Conti et al. [94] proposed two different algorithms for counting people in a classroom, both based on CNN. The results show that they can reach very good people-counting effects. Early research focused on estimating the number of people by detecting the body or head, and some were based on the mapping of local or global features to the actual number. Recently, the population-counting problem has been expressed as a regression of a population density image, with the number of people in the image then obtained by summing the values of the density image. This method can handle severe occlusion in dense crowd images. With the success of deep learning technology, researchers used CNN to generate accurate population density images and achieved better performance than traditional methods. Some scholars have applied CNN to the field of crowd counting in complex scenes and achieved good results [53,106].

3.3.4. KNN

KNN was first proposed by Cover and Hart in 1968. It is a theoretically mature method and one of the simplest machine learning algorithms. The idea of this method is very simple and intuitive: If most of the nearest samples in the feature space of a sample belong to a certain category, the sample also belongs to this category and has the characteristics of the samples in this category. In determining the classification decision, this method only determines the category of the sample to be classified based on the category of the nearest one or several samples [71]. KNN’s method is simple, easy to understand, easy to implement, and does not need to estimate parameters. Because the KNN method mainly depends on limited surrounding nearby samples, rather than the method of discriminating class domains, the KNN method is more suitable for the sample set to be divided with more cross or overlapping class domains. One of the shortcomings of this method is the large amount of calculation, because for each sample to be classified, the distance to all known samples must first be calculated to find its nearest neighbors. In a research work by Vela et al. [52], KNN was selected as the best algorithm to estimate occupancy. But because of its O(n2) time complexity, its use needs to be considered. This algorithm also has a shortcoming in classification. When the sample is unbalanced, when, for example, the sample size of one class is very large, while the sample size of other classes is very small, this may cause, when a new sample is input, confusion among the K neighbors of the sample when large-volume samples account for the majority. Szczurek et al. [78] verified that the nonparametric, nonlinear, minimum distance classifier i.e., KNN is very effective in occupancy determination.

3.3.5. Regression

The regression algorithm is the most common and widely used in machine learning. The most commonly used regression algorithms are linear regression and logistic regression. The definition of linear regression is that the target value is expected to be a linear combination of input variables. The linear model is simple in form and easy to model, but it contains some important basic ideas in machine learning. Linear regression is widely used. It is a statistical analysis method that uses regression analysis in mathematical statistics to determine the quantitative relationship between two or more variables. Kim et al. [72] used a linear regression analysis to find the correlation between occupancy and electricity consumption. Logistic regression models the relationship between the feature and the target as a logistic function, predicting the possibility that the new test sample belongs to a certain category, and then uses the maximum likelihood estimation method to estimate the parameters. Regression analysis studies the relationship between one or more independent variables with one independent variable. This method can be used in determining the relationship between occupancy rate and the concentration of the captured property of the sensor, and so on. Softmax Regression Model is a generalization of logistic regression in multi-classification. In Yuan et al. [48], the Softmax Regression Model is used to calculate the emission probability matrix for clarifying the dynamic relationship between environmental parameters and indoor occupancy. Mohottige et al. [66] employ a regression stage to better relate the output of the classifier to the actual room occupancy.

4. Discussion

4.1. Choices in Estimating Occupancy Counting

A large number of researchers have developed a variety of methods for estimating the number of people in a given room, and tested the feasibility of the method through experiments. It is worth noting that these approaches differ in the collocation and selection of occupancy measurement systems and algorithms. In practical applications, how to choose the method of estimating occupancy counts can be considered from the following four aspects according to the actual situation: range of people, accuracy, cost, and privacy.

4.1.1. Range of People

Different occupancy measurement systems have their corresponding advantages, and choosing to use them according to their advantages can maximize their effectiveness. Sometimes we need to detect the occupancy of a larger environment, such as a classroom; sometimes we only need to detect occupancy in a small room with a relatively small number of people to monitor. How to choose the right occupancy measurement system for the different numbers of people range is our concern. In this review, the number of people to whom the literature reviewed was applied was counted, and the range of people was divided according to the following criteria: more than twenty-five people were classified as a high-density population, five to twenty-five people as a medium-density population, and less than five people as a low-density population. The literature screened for this review was statistically summarized according to the above criteria. A statistical plot of the number of different occupancy measurement systems used in experiments with the different number of people ranges was drawn, as shown in Figure 7.

According to Figure 7, it is obvious that using infrared for occupancy counting in a small range of people is more popular, and almost half of the researchers choose to use infrared for occupancy counting experiments. Because of its advantages, such as low cost, easy installation, and better privacy, the use of infrared for occupancy counting can be preferred in small-area headcounts without special requirements. In the medium-density population, on the other hand, the preference for a certain measurement system selection does not point in a particularly clear direction and can be chosen according to actual needs and one’s capabilities. In a high-density population, the choice of measurement system will face more problems. Mainly, the environment in which the high-density population is located is relatively complex, and many sensors are difficult to apply in high-density situations. The requirements for the experimenter and the experimental environment will be higher, and it is difficult to achieve excellent estimates of occupancy counts. Among the reviewed literature, occupancy counts based on communication technology is the most commonly used in high-density populations. It is relatively less dependent on the environment, and although the inherent flaws of wireless communication technology cannot be avoided, it can achieve relatively good results. The use of wireless communication technology can be preferred for high-density population counting.

In summary, when only considering the measurement needs of different densities of occupants, in pursuit of better cost performance, researchers prefer to use infrared to conduct experiments in the estimation of low-density population occupancy, and the occupancy counting method based on wireless communication technology is adopted in high-density crowds. However, in the case of medium-density populations, there is no obvious preference for the use of the occupancy measurement system, and other selection criteria need to be comprehensively considered.

4.1.2. System Usage Accuracy

The accuracy of the system is defined as follows: the occupancy counting system can accurately determine the number of residents in a given location within the error range. The level of accuracy required by the system may also vary depending on specific needs. For example, for HVAC applications in smart buildings, as long as the overall occupancy trend can be reflected, certain estimation errors can be tolerated, while for emergency evacuation applications in the same building, 99% accuracy may be required. Occupancy counting accuracy can be defined as how close the measured value is to the actual occupancy on the ground. In terms of accuracy, both occupancy measurement systems and occupancy counting algorithms have an impact on accuracy. These two aspects will be discussed next, which can give reference to hardware and software accuracy selection.

In this study, box plots are used to show the distribution of the accuracy data. A box plot is a statistical chart used to show the dispersion of a set of data. It is named because of its box-like shape. Box plots provide a visual representation of the data distribution characteristics. When the box plot is short, it means that many of the data points are similar because many of the values are distributed over a very small range. When the box plot is high, it means that most of the data points are very different from each other because the values are widely distributed. If the median is closer to the bottom, then most of the data have lower values. If the median is closer to the top, then most of the data have higher values. Basically, if the median line is not in the middle of the box, then the plot indicates skewed data. A long line at the top and bottom of the box indicates that the data has a high standard deviation and variance, meaning that the values are spread out and vary to a high degree. If there are long lines on one side of the box and not on the other side, then the data may vary to a high degree in only one direction. The box plot shows very rich information for the next step of the analysis. The biggest advantage of the box plot is that it is not influenced by outliers and can accurately and stably depict the discrete distribution of the data, as well as facilitate data cleaning.

A study [15] compared four different methods of estimating occupancy: (i) overhead video-based occupancy counting system, (ii) PTZ camera face detection system, (iii) CO₂⁻ based physical model, (iv) CO₂⁻ based statistical model. The experimental results showed that the PTZ-camera face-detection algorithm was significantly better than the other three methods in terms of accuracy. The comparison between the four showed that the selection of different sensors and models under the same experimental environment resulted in differences in accuracy. In a single type of measurement system, the difference in accuracy, in addition to the measurement system itself, produces errors, but also, with the external environment, changes in the exogenous variables have a very important relationship with the calculations. Therefore many researchers have conducted repeated multiple experiments in different places to test the performance of the proposed method. The highest and lowest accuracy values of the experiments they performed are listed in the accuracy column of Table 1. Since the experiments depend on large randomness of changes such as environment, some of the anomalous data are excluded for discussion based on the actual situation. By analyzing the experimental accuracy of the occupancy measurement system, the accuracy box plot is drawn as shown in Figure 8. The box plot shows the accuracy maximum value, minimum value, median, average value, etc. This method of data display is very convenient for comparing and analyzing each variable.

The historical best accuracy of various occupancy measurement systems is above 95%, which shows that under certain conditions, various sensors have suitable occasions for use, and better accuracy can be obtained. Therefore, it is very important to choose the right sensor according to the situation. From the height of the box, it can be seen that the environmental sensors have a wide range of measurement accuracy, indicating that their measurement accuracy distribution is fluctuating within a certain range and the accuracy varies greatly. This places higher demands on the selection and operation of the experimenter. Because environmental sensors naturally have a relatively large disadvantage, namely, their dependence on the environment, since the environment is something that can change easily. Therefore, it is a prerequisite for achieving a good result while using environmental sensors to fully consider the impact of changing experimental environmental conditions and make adjustments accordingly. The accuracy data of the occupancy measurement system using the camera and environmental sensors have high variance, indicating that their measurement accuracy is not very stable. In the camera-based box plot, we see that the median line is clearly located in the upper half of the box, indicating that the accuracy of the camera-based experiment is good in most cases. Generally, image/video-based methods have high accuracy, but the accuracy of camera experiments fluctuates greatly, and the stability of the accuracy is poor, because it is greatly affected by occlusion, uneven density, changes in scale and viewing angle. Choi et al. [28] summarized the main cases causing accuracy bias after performing occupancy calculations based on vision, which is more informative for later researchers. Lighting conditions change rapidly (when on or off) and misrecognition of things such as chairs and winter clothes can lead to overcounting. The following three conditions can lead to undercounting: (1) The occupant is far from the camera (more than 10 m), and the partition is obscured. (2) The occupant is close to the camera but only the back of his/her head and chair are visible. (3) The occupant is close to the camera but is obscured by a partition or person. Relatively speaking, the accuracy fluctuation of the experiment using wireless communication equipment is the smallest, and its scalability is strong. It is suitable for occasions that require high accuracy and stability. The camera-based and infrared occupancy measurement outliers are larger, so to achieve better experimental accuracy, one would need to have better control over the installation, environment, hardware selection, etc. Otherwise, those methods will also lead to large differences in the same environment. In response to the accuracy problem of a single sensor, some researchers have also targeted the use of multiple sensors in conjunction with each other to improve accuracy. Candanedo et al. [79] evaluated the accuracy of HMM using different environmental parameters (temperature, humidity, humidity ratio, CO₂ and light time series data) for occupancy estimation. Data collection and data fusion of different environmental parameters showed different accuracy. Sensor fusion also helps to improve accuracy. Longo et al. [60] compared the occupancy accuracy of different scenarios using all features, Wi-Fi-only, or Bluetooth/BLE-only features. It is clear that the accuracy is significantly improved after fusion.

Each sensor has specific advantages and disadvantages as far as occupancy count estimation is concerned. In terms of measurement accuracy, to determine the number of occupants in a room, the camera has high detection accuracy and is suitable for applications that require accuracy in occupancy counts. Environmental sensors are affected by the external environment, and accuracy measurement fluctuations are relatively large, resulting in the need to control the impact of extraneous variables to achieve better results. Due to the limitations of environmental sensors and infrared, it is recommended that this type of sensor be used when the accuracy requirements are not very high, and only the approximate numerical range of in the room needs to be estimated to reduce the difficulty of monitoring. The accuracy fluctuation of the wireless communication equipment experiment is the smallest, and the methodology’s performance is relatively stable. It is suitable for occasions that require high accuracy and stability. Combining multiple sensors is a good choice to improve the estimation accuracy, and the specific matching method needs to be selected according to the experimental situation.

4.1.3. Algorithms Usage Accuracy

Similarly, the choice of occupancy-counting algorithms also has a great impact on accuracy. Wang et al. [71] compared three popular machine learning algorithms, including KNN, SVM, and ANN, combined with three data sources, namely, environmental data, Wi-Fi data, and fused data. Through field experiments, the performance of different algorithms and different data sets are compared as to reliability and accuracy. Fusing multiple sensing data sources was helpful in improving the reliability and accuracy of the ANN model. Different algorithms showed different performances for the same experimental environment. To compare the effects of different algorithms on the accuracy of occupancy counting, the five most frequently used machine learning algorithms discussed in Section 3.3 are analyzed.

The accuracy data from literature where these algorithms were used were taken for plotting, and the accuracy box plots were drawn as shown in Figure 9. It is obvious that different algorithms have different accuracy levels and stability. Choosing an appropriate algorithm with the occupancy measurement system will select different results. It is necessary to comprehensively consider the accuracy requirements, stability, algorithm complexity, compatibility and other aspects, and reasonably select the algorithm. According to the box plot distribution, we can see that the various types of algorithms achieve relatively good results in most cases, but that there are some biases. The variance of the SVM algorithm is the largest and the accuracy fluctuates considerably. The average accuracy of the KNN and regression algorithms is relatively high compared to the other algorithms. The ANN and KNN vary less in the high accuracy direction and the regression algorithm varies less in the low accuracy direction. The performance of the CNN algorithm is more moderate. SVM has excellent performance on many datasets. Relatively speaking, the nature of SVM is to try to maintain the distance from the sample, which leads it to be more resistant to attacks. It is an algorithm that you can try first when you get the data. ANN can learn and construct models of complex nonlinear relationships. The algorithm can better model heteroskedasticity, i.e., data with high volatility and unstable variance, because it can learn hidden relationships in the data without imposing any fixed relationships in the data. CNN is often used for camera-based occupancy counting, and it has great advantages in processing images. It can effectively reduce the dimensionality of large data volume into small data volume, and it can effectively preserve the image features following the principle of image processing. KNN can be used for both classification and regression, and has a lower training time complexity than algorithms such as SVM. This algorithm is more suitable for the automatic classification of class domains with larger sample sizes, while those with smaller sample sizes are more prone to misclassification cases when using this algorithm. The regression algorithm is mainly used for occupancy estimation through relational fitting and is suitable for applications where the inputs and outputs are significantly related. Careful design based on some prior knowledge of the data is required to select the best parameters. There may be some limitations in the use of individual algorithms, and by cascading algorithms the advantages of each type of algorithm can be combined to improve the accuracy of occupancy estimation. Zou et al. [81] proposed a new cascading video analysis algorithm based on a creative combination of SVM, CNN, and K-mean clusters. It successfully outperforms these individual algorithms in terms of accuracy and computational cost. The experiments by Singh et al. [61] demonstrate that algorithm selection is very important, but with careful selection, it can obtain up to 100% accuracy when detecting user presence. In addition, they prove that sensor placement plays a crucial role in system performance and that the best results are obtained by placing the sensors on the ceiling of the room. In addition, the detection accuracy can also be improved by eliminating the noise in the data. Kim et al. [107] introduced the label noise filtering method to eliminate the suspicious noise data in the data. Through comparison, it is found that only by eliminating the noise, the detection accuracy is greatly improved.

SVM has excellent performance on many datasets, so you can first try to use the SVM algorithm as a control. CNN is often used in image processing and has a greater advantage, so camera-based occupancy counting can be preferred to use this algorithm. KNN algorithm is suitable for the occasion of relatively large sample capacity. ANN can deal with situations where the nonlinear relationship is more complex. The regression algorithm is mainly used for occupancy estimation through relationship fitting, and is mainly used for occasions where the input and output have obvious relationships. There may be some limitations in the use of a single algorithm. The cascade algorithm can be considered to integrate the advantages of various algorithms to improve the accuracy of occupancy estimation. Filtering and eliminating noise in the data before using various algorithms to process the data can also be a good way to improve detection accuracy.

4.1.4. Cost and Privacy

As for cost and privacy, it is necessary to select sensors that meet the requirements based on the characteristics of various occupancy measurement systems. Some existing occupancy estimation systems can achieve impressive accuracy, but they either require a labor-intensive calibration phase or installation of bespoke hardware such as CCTV cameras, which are presumptively privacy-intrusive.

Cost is divided into hardware cost and software cost. To facilitate adoption in different application scenarios, the occupancy counting system should bear the lowest hardware, installation, and maintenance costs. This may require the use of existing infrastructure or the use of off-the-shelf hardware, or it may require the development of innovative hardware components to easily adapt to existing infrastructure delays. To be acceptable to participants in the crowd, from a privacy perspective, the people-counting system should be non-invasive. In addition, it should only collect the information needed to provide the service, and ideally, it should be difficult to use this information in irrelevant places. If there is no high requirement for accuracy, the indirect method based on environmental sensors will be more popular. From the user’s point of view, there are two reasons for choosing the indirect method. The first reason is that the price is relatively inexpensive, and the second reason is that these sensors will not infringe on the privacy of the occupants. Although the accuracy of camera-based systems is relatively high, the cost is often relatively high, and it will face privacy violations.

4.2. Challenges

Although there has been a lot of research on estimating occupancy counts, and the corresponding technology has been updated and developed, the current research in this area is still full of many challenges.

4.2.1. Challenges of High-Density Crowd

The traditional method is limited to low-density groups of people. As the population grows, political gatherings, sporting events, concerts, and other gatherings become more frequent, so the need for high-density crowd analysis is also increasing. But crowd segmentation is a challenging problem, which means that counts cannot be accurately obtained in most crowded scenes. In actual application scenarios, there may be very dense crowds and severe occlusion. When two or more objects are very close to each other and merged, occlusion will occur. Only a part of the human body can be shown in the image, and it is difficult to identify a single object, resulting in reduced counting accuracy. There may also be a stationary crowd at the scene. In a high-density crowd, it is easy to mix with others and cause miscounting. Few systems can achieve good crowd counting effects in high-density crowds. Many studies are aimed at low-density crowd counting, even limited to a few people.

4.2.2. Challenges of External Factors

The influence of external factors such as the external environment, sunlight, weather, etc. also brought great difficulties to counting the number of people. In terms of environmental sensors, the sensors are significantly affected by external factors, and any small environmental change may have an impact on the final counting effect. In daily life, simply opening windows to ventilate will cause earth-shaking changes in the indoor environment. To avoid such effects, some researchers using environmental sensors have chosen to conduct system tests in a strictly airtight and unventilated indoor environment. However, the strict requirements for the experimental environment mean that the scope of application of the system will be greatly reduced.

The counting effect of camera-based sensors is also affected by external factors. Affected by factors such as shooting angle, lighting, imaging distance, people’s posture, clothing and accessories (such as scarves and hats), the appearance of different people is ever-changing, and the differences are very large. This brings a huge challenge to the human body recognition classifier. Due to different camera angles, tilts, and up and down movement of the camera position, the angle of view will change. Object recognition and counting accuracy are greatly affected by different viewing angles [108]. The complex background of the scene is also a big challenge. Whether it is indoors or outdoors, the detection of human bodies generally operates against very complex backgrounds. Some objects are very similar in appearance, shape, color, and texture to human bodies. As a result, the algorithm cannot accurately distinguish the human body from the background of the scene.

4.2.3. Challenges of System Robustness

Although important for practical applications, few studies focus on cross-scene crowd counting. It seems that no sensor or model can be best installed in all institutional buildings. Most systems are suitable for the same one or a class of scenes, and can achieve higher accuracy only under fixed requirements. The crowd counting model learned for a specific scene can only be applied to the same scene. Given a new scene or changing the layout of the scene, the model must be retrained according to the new features of the scene. This will limit the use of the system. If the system is to be put into use better, more consideration should be given to the robustness of the system. Many authors have researched the robustness of the system. Sharma et al. [44] have explored using the same CNN model in a different home setting with a single occupant. High counting accuracy can still be maintained under placement variations of tags, receivers, and furnishing. There are many similar studies, but it can be clearly seen that these studies still have limitations. The test conditions they usually set are relatively simple. In the face of a more complex and changeable environment, the performance of the system needs to be considered. Whether the system can maintain stability over time is also an important problem. The current studies on occupancy counting have one of the most obvious flaws. There are not enough studies using long-term measurement data to evaluate whether occupancy counting can maintain accuracy under seasonal changes.

5. Conclusions and Future Directions

In this paper, we review studies from the last ten years, focusing on the literature on estimating occupancy counts. Through the analysis of the distribution characteristics of the literature, it is found that the enthusiasm of researchers in various countries for estimating occupancy counts is high, especially after 2016. Researchers from various countries have undertaken a significant amount of exploration and efforts in these areas and have obtained more research results. Developing countries (headed by China) and developed countries (headed by the USA) have carried out a large number of research projects. For many occupancy measurement systems based on different sensors, the advantages and limitations of different sensors are discussed. We have classified the systems according to the different sensors used, such as environmental sensors, camera, infrared, wireless communication, etc. The fusion of multiple sensors is also reviewed. When selecting sensors for practical applications, sensor fusion methods tend to show better performance, because the advantages and disadvantages of different sensors can compensate for each other. We reviewed the commonly used occupancy counting algorithms and found that machine learning algorithms are the most popular, especially those based on supervised learning.

As for how to choose occupancy measurement systems and occupancy counting algorithms in practice, one should consider the four aspects: range of people, accuracy, cost, and privacy. When only considering the measurement needs of coping with different density occupancy counts, preference can be given to infrared for occupancy counting when facing low-density crowds, while in high-density crowds, the occupancy counting method based on wireless communication technology is used. In terms of accuracy, the selection of either occupancy measurement systems or occupancy counting algorithms will have an impact on accuracy. The use of environmental sensors is more dependent on the environment, and full consideration of the impact of changes in experimental environmental conditions is a prerequisite for achieving a good result with environmental sensors. Environmental sensors and infrared have many limitations. It is recommended that such sensors be used when the accuracy requirements are not very high, and only the approximate range of people in the room needs to be estimated. Generally, image/video-based methods have high accuracy, but many factors affect the accuracy of recognition and counting, the accuracy of the experiment fluctuates greatly, and the accuracy of stability is poor. The accuracy fluctuation of experiments using wireless communication equipment is the smallest, and its scalability is strong. It is suitable for occasions requiring high accuracy and stability. Sensor fusion is a good option to improve estimation accuracy. Different algorithms have different levels of accuracy and stability. Choosing an appropriate algorithm with the occupancy measurement system will get different results. Among the commonly used occupancy estimation algorithms, the SVM algorithm has excellent performance on many datasets, and oner can first try to use the SVM algorithm for comparison. CNN has great advantages in image processing, and CNN can be prioritized for camera-based occupancy counting. The KNN algorithm is suitable for occasions with relatively large sample sizes. ANN can handle complex situations with nonlinear relationships. The regression algorithm is mainly used in situations where there is a clear relationship between input and output. Filtering and eliminating the noise in the data can also improve detection accuracy. There may be some limitations in the use of a single algorithm. The cascade algorithm can be considered to integrate the advantages of various algorithms to improve the accuracy of occupancy estimation.

Although the corresponding technology has been updated and developed, occupancy counting still faces many problems and challenges to be solved in practical applications. There are still many pressing issues in terms of high-density crowd occupancy counting, complex environmental impact, and system robustness. Regarding future trends, with the advent of the 5G era and the development of network technology, the data transmission rate of sensors will be greatly accelerated, and the response speed of occupancy measurement systems will also increase. The advent of the big data era has brought better technical support for data conversion, data processing and data storage. With the development of technology in the new era, there will be more room for development and progress in the study of occupancy counting In recent years, cloud-based data processing has become an important aspect of computations [109], and deep learning methods have gradually been used for occupancy measurement. Therefore, the rapid development of cloud computing platforms and artificial intelligence, especially deep learning technology, will have a positive impact on this field. In the past, some sensors were used to count people indoors, such as Bluetooth, Wi-Fi, RFID, infrared, and CO₂ concentration sensors. In recent years, as deep learning has become popular, indoor people-counting methods based on computer vision have become more and more popular. With the rapid development of cloud computing platforms and artificial intelligence, especially deep learning technology, there will be more in-depth research and breakthroughs in new system models and algorithms in the future. It can be seen from the extensive literature in this field that the research community pays significant attention to building occupancy estimation. But there are still many research issues to be solved. In the future, further research may focus on the following aspects:

Many occupancy measurement systems are affected by changes in the external environment. It is possible to try to realize the optimization method of real-time occupancy estimation in the embedded system, predict the changes in the surrounding environment in advance and continuously improve the accuracy of the model through automatic correction and adjustment. New, more effective, robust, reliable and efficient estimation models can continue to be developed, assessing the impact of each variable on the empirical model and balancing model performance with practical feasibility. In terms of algorithm research, with the gradual introduction and development of new technologies, the innovation and optimization of occupancy estimation algorithms can be a direction for continued research, especially in emerging machine learning. Effective algorithm research can be used to improve estimation accuracy and solve problems such as effective segmentation of high-density populations. The universality of the occupancy system needs to be improved. There is currently no single sensor or model that works best for all buildings, making it difficult to generalize the relevant concepts to every building. Simpler and easier-to-apply methods might be sought through future research work. In terms of application, estimating the occupancy count can be combined with intelligent control devices applied in building construction, which can effectively solve the problem of excessive energy consumption and create a smart city. This review can provide a reference for subsequent researchers to conduct occupancy counting research, and give suggestions on the selection of occupancy measurement systems and algorithms, as well as understand the technology development trends in recent years, the limitations of current research and future development directions. It can be used by other researchers as a starting point for their research work.