Quality of Automatic Traffic Volume Counting by Cameras and Impact on the Qualitative Indicators of Traffic

Abstract

This paper presents research on the collection, analysis, and evaluation of the fundamental data needed for road traffic systems. The basis for the research, analysis, planning and projections for traffic systems are traffic counts and data collection related to traffic volume and type. The quality and accuracy of this data are very important for traffic planning or optimization. Therefore, the purpose of this research is to apply advanced methods of automatic counting of motorized traffic and to evaluate the impact of this data on the measurement of important traffic indicators. The accuracy of measurements arising from the traditional method of data collection through manual counting will be compared with the most advanced methods of automatic counting through cameras. For this purpose, an analytical algorithm for the recognition and processing of data related to road users as a function of the time of day was applied. The program was written in the programming language Python, and the accuracy of the data and its effect on the results of qualitative traffic indicators were analyzed using the Synchro software model. The developed program is capable of recognizing and classifying different types of vehicles in traffic, such as motorbikes, motorcycles, cars, pick-ups, trucks, vans and buses, as well as counting the traffic volume over time. The results obtained from these two models show the advantages of applying advanced methods of data collection and processing related to dynamic traffic processes, as well as the quality in terms of the impact on the measurement of qualitative traffic indicators. A comparison of the quality of results for the different time intervals and varying levels of visibility in traffic is presented using tables and graphs. At nighttime, when visibility was poor, the discrepancy between the manual and automatic counting methods was around 9.5%. However, when visibility was good, the difference between manual counting and the automated program was 4.87% for the period 19:00–19:15 and 3.64% for the period 05:00–05:15. This discrepancy was especially noticeable when distinguishing between vehicle categories, due to the limitations in the accuracy in recognizing and measuring the dimensions of these vehicles. The difference between the two calculation models has a minor effect on qualitative traffic indicators such as: approach LOS, progression factor, v/s, v/c, clearance time, lane group flow, adj. flow, satd, and flow approach delay.

1. Introduction

Planning a traffic road network, designing traffic systems, or modelling these systems in order to manage traffic problems starts with collecting data about the traffic volume or basic data related to the road users. The demand for motorized traffic has increased enormously in urban centers and this is accompanied by major problems in traffic operation being manifested in a decrease in the quality of motorized and unmotorized movement, traffic delays and other safety problems [1]. The main indicators that are related to traffic demand and are important for modelling, optimization, and the analysis of proposed solutions are based on these traffic measurements. In this regard, the quality and accuracy of these basic data are very important for the other processes of traffic planning or optimization to show acceptable results. In relation to the increased demand for motorized movement and the need for temporal and spatial non-linearity analysis, there is a need to find a new approach or a new, more favorable measurement model such as the automatic numbering model, which would save time compared to manual counting and the results would be qualitative. Due to the impact on the measurement of qualitative indicators of traffic, especially of heavy vehicles in traffic, the accuracy of the categorization of the vehicles that are part of motorized traffic is very important. The objective of the study is to enhance the accuracy of qualitative traffic indicators by exploring the effect of the techniques employed for qualitative baseline measurements.

Consequently, we will answer three questions:

• What is the difference in data accuracy between the manual and automatic numbering models?
• What is the impact of the basic data obtained from two models in the determination of qualitative traffic indicators?
• How accurate is the data obtained by automatic counting in conditions of poor visibility at night?

We address these questions below, with the concrete results obtained by also applying the data comparison model. Traffic processes are complex and dynamic; therefore, there is a need for frequent traffic counts in relation to time and space, especially in urban areas, because there is a need to adapt the forms of traffic control, traffic optimization or supply, in accordance with the traffic’s demand [2]. Until recently, traffic volume counts were usually realized through a manual method. Now, with the beginning of the use of advanced automatic techniques, special devices are being applied through which the realization of traffic counts is enabled automatically. Therefore, the analysis and evaluation of the quality of the basic information (traffic counts) obtained through cameras and the processing of this information in comparison to the methods of manual traffic counting have been the target of this research. Opportunities to bring engineering support specialties together through technology may be found by utilizing the performance technology and database capabilities of technology information management systems [3]. YOLO has developed a similar system which consists of dividing an image into N equally sized grids of M × M. Each grid is responsible for detecting and localizing any objects that are present in its region [4]. These grids can also predict the bounding box coordinates of an object, relative to the coordinates of the cell, as well as the object marker and the probability of the object being present in the cell. This process reduces the computational burden since both the detection and recognition of objects are handled by the cells of the image; however, it can lead to a lot of repetitive predictions due to numerous cells detecting the same object with distinct bounding box predictions [4]. Some authors [5] have proposed a scheme as part of a research project that concluded shortly before the public release of YOLO version 4, which includes new methods that likely improve classification, detection, and counting rates [5]. Abdelwahab [6] evaluated the proposed method; three experiments were employed using five videos representing different circumstances. In all experiments, the GMM was used for creating the background model for only the ROI [6]. Other authors (Ref. Lin Zheng, Junjiao) provided a new method for the rapid detection and classification of traffic anomalies and improved the accuracy of detection and classification [7]. Although manual methods for traffic counting have been shown to be successful in terms of the quality of the data obtained, they require time and commitment, a large number of staff, and also the data processing requires time equivalent to the counting time. Therefore, the application of automatic counting methods, enabling efficiency in terms of the time required to obtain the information and the data processing as well as the quality of this information, is an enormous contribution to the planning, projection and efficiency of traffic management. Automated traffic data collection has been sought in transportation applications to reduce cost and improve efficiency compared to manual data collection [8].

Approach and Methodology

For research purposes, an algorithmic model was applied to recognize the different categories of vehicles in traffic, enabling data processing. The program was built using the programming language “Python” and the accuracy of the data was analyzed through the built-in software model “Synchro” to assess the impact on the results of qualitative traffic indicators. Object detection is the process of predicting the class of one or more objects within an image and drawing a bounding box around them (Figure 1). This is commonly used to identify specific objects, such as vehicles, in a given frame [9]. Deep learning is a form of artificial intelligence where the system tries to learn from data without the need for labels or pre-defined features. In particular, convolutional neural networks (CNNs) are used to perform unsupervised object detection [10].

Figure 1. Object detection process of predicting the class of one or more objects within an image.

Object detection models generally have two corridors:

• An encoder takes a frame (image) as input and applies a series of layers and blocks to it in order to generate statistical data that can be used to identify and label the objects within the frame, such as vehicles;
• The encoder will send the data to a decoder, which will then use the information to generate bounding boxes and labels for every object present [11,12,13].

The tracker of objects utilizes the Euclidean distance concept to monitor an object [4].

The authors of [14] formulated a speedy algorithm that can be used to count vehicles in traffic videos without requiring vehicle tracking. A reference model was generated for a small area in the video frame to enable this and the proposed algorithm increased the speed of video processing by utilizing every third frame instead of every frame [14].

The authors of [15] implemented a virtual detection line in each of the traffic lanes to count the number of cars. The points at which the vehicles entirely crossed these lines were determined by these lines [15]. In this research, they offered an alternative and new approach for counting the number of vehicles crossing the road in video sequences by recognizing motion using incremental subspace learning [15]. YOLOv4 is an integrated algorithm combining detection and recognition, which can directly obtain the location and category of the target from an image. This algorithm is an improved version of YOLOv1, YOLOv2, and YOLOv3 [16]. One of the primary innovations of the work by the authors [17] is the edge-based technique for preprocessing and filling missing sensor data, which takes into consideration both temporal and geographical information. It achieves better accuracy in handling small targets due to the addition of CSPNet in the CNN design of YOLOv4. YOLOv4-Tiny is a simplified version of YOLOv4, which reduces the accuracy compared to YOLOv4. This is because the YOLOv4-Tiny backbone network is relatively shallow and is unable to extract higher-level semantic features [18]. In order to meet the requirements of high detection speed and accuracy, the proposed target detection algorithm improved YOLOv4-Tiny. YOLOv4 divides an image into an M × M grid before it is fed into the neural network, with each grid cell responsible for predicting objects. Each grid is capable of predicting up to B bounding boxes, each of which has an associated confidence score [19]. Each box contains five variables which are defined in Equation (1).

$$P=[C_1,C_1,\cdots ,C_n]$$

(1)

$$P_{Otensor}=\mathbb{N}·\mathbb{N}·\left(Numbe{r}_{o{f}_{pixels}}·\mathbb{Q}+P\right)$$

(2)

$$\mathscr{H}={\displaystyle \sum }_{\begin{array}{c}i=0\\ j=0 \end{array}}P{\left(i,j\right)}^{\mathbb{N}, \mathbb{Q}}\ast \left[\acute{\Re }lo{g}_{\left(\acute{\Re }\right)}+lo{g}_{\left(\acute{\Re }\right)}\left(1-{\acute{\Re }}_i\right)\left(1-\Re \right)\right]-{\displaystyle \sum }_{\begin{array}{c}i=0\\ j=0 \end{array}}nob{\left(i,j\right)}^{\mathbb{N}, \mathbb{Q}}\ast \left[P_{Otensor}lo{g}_{\left(\acute{\Re }\right)}+log\left(1-{\acute{\Re }}_i\right)\ast \log \left(1-P_i\left(o\right)\right]\right]$$

(3)

2. Model Design and Automatic Measurements through Cameras

For the needs of the model design and to achieve the goal of this research, traffic counts were made in the road segments between the 10 intersections of the national road N9 in Fushe Kosove, Figure 2.

Figure 2. Road network maps where traffic counts were made.

The counts were realized within a time interval of 24 h, applying the methods of manual and automatic counting through cameras. Then, a comparison of the data obtained through these two methods was made in terms of the quality of this data and the impact they may have on the application of other important models for the management of traffic problems. Figure 3 shows the flow chart of the methodological approach for the traffic count and qualitative indicator measurement.

Figure 3. Process and methodology used for the traffic count.

Parts of the analytical pseudo Algorithms 1 and 2 are shown below.

Algorithm 1: Finding the left of the frame (image).

Input:$\mathrm{Parameters}x, y h and w$

Output:$\mathrm{Banderas}\mathrm{of}\mathrm{the}\mathrm{different}\mathrm{vehicles},N_1\to N_N$

$\mathit{i}\mathit{n}\mathit{t} \in \left(x, y,h and w\right)$

$\mathit{f}\mathit{o}\mathit{r} 1 to N \mathit{d}\mathit{o}$

$\mathit{s}\mathit{e}\mathit{l}{\mathit{f}}_{\left(\mathit{x},\mathit{y}\right)} \to w, h$

$\mathit{s}\mathit{e}\mathit{l}\mathit{f}{\mathit{c}}_{\mathit{x}} \to x+\mathit{s}\mathit{e}\mathit{l}{\mathit{f}}_{\left(\mathit{x},\mathit{y}\right)}$

$\mathit{s}\mathit{e}\mathit{l}\mathit{f}{\mathit{c}}_{\mathit{y}} \to y+\mathit{s}\mathit{e}\mathit{l}{\mathit{f}}_{\left(\mathit{x},\mathit{y}\right)}$

$\mathit{r}\mathit{e}\mathit{t}\mathit{u}\mathit{r}\mathit{n}\to \mathit{s}\mathit{e}\mathit{l}{\mathit{f}}_{\left(\mathit{x},\mathit{y}\right)}$

$\mathit{d}\mathit{e}\mathit{f}\to$$\mathrm{different}\mathrm{vehicles},N_1\to N_N\ast \mathit{s}\mathit{e}\mathit{l}{\mathit{f}}_{\left(\mathit{x},\mathit{y}\right)}\left[\mathit{N}\ast \mathit{M}\right]$

$\left(x, y,h and w\right)\cong I_d \exists I_{d\to box}$

$\mathit{C}\mathit{e}\mathit{n}\mathit{t}\mathit{e}\mathit{r} \forall \mathit{s}\mathit{e}\mathit{l}{\mathit{f}}_{\left(\mathit{x}, \mathit{y}, \mathit{h} \mathit{a}\mathit{n}\mathit{d} \mathit{w}\right)}$

$\mathit{f}\mathit{o}\mathit{r} x_i,y_j to Center \mathit{d}\mathit{o}$

# Finding the position of the vehicle

$\mathit{i}\mathit{f} y_j\to P{o}_{line}$$\mathrm{and}{y}_j\to P{o}_{midel\_line}$

$\mathit{i}\mathit{f} \nexists {\mathit{U}}_{\mathit{l}\mathit{i}\mathit{s}\mathit{t}}$ and $U_p \exists U{p}_{list}$

$\mathit{e}\mathit{l}\mathit{i}\mathit{f} y_j \exists D_{o{w}_{p{o}_{line}}}$$\mathrm{and}{y}_j \exists D_{o{w}_{mide{l}_{line}}}$

$\mathit{e}\mathit{n}\mathit{d}$

$\mathit{i}\mathit{f} \nexists D_{o{w}_{po\_line}}\to I_d \exists I_{d\to box}$

$\mathit{e}\mathit{l}\mathit{i}\mathit{f} y_j>D_{o{w}_{po\_line}}$

$\mathit{i}\mathit{f} \nexists U{p}_{list}\to$$I_d \exists I_{d\to box}$

$\mathit{e}\mathit{l}\mathit{i}\mathit{f}$$y_j>D_{o{w}_{po\_line}}$

$\mathit{e}\mathit{n}\mathit{d}$

${\mathit{U}}_{\mathit{l}\mathit{i}\mathit{s}\mathit{t}} \cong I_d \exists I_{d\to box}$

$\mathit{C}\mathit{e}\mathit{n}\mathit{t}\mathit{e}\mathit{r} \forall \mathit{s}\mathit{e}\mathit{l}{\mathit{f}}_{\left(\mathit{x}, \mathit{y}, \mathit{h} \mathit{a}\mathit{n}\mathit{d} \mathit{w}\right)} \to \mathbf{2}\mathit{R}\mathit{G}\mathit{B},\pm \mathbf{1}$

$\mathit{e}\mathit{n}\mathit{d} \mathit{e}\mathit{n}\mathit{d}$

Algorithm 2: Part of the algorithm for counting and categorizing vehicles during automatic counting through cameras.

import cv2

class RealTime(Processing):

def __init__(self):

super().__init__()

def realTime(self):

# Set the frame rate of the video capture device

cap = cv2.VideoCapture(0)

cap.set(cv2.CAP_PROP_FPS, 800)

while True:

# Read the video frame and resize it

ret, frame = cap.read()

frame = cv2.resize(frame, (self.width, self.height))

# Create a blob from the image

blob = cv2.dnn.blobFromImage(frame, self.scale, (self.width, self.height), (self.mean,                        self.mean, self.mean), swapRB=True, crop=False)

# Set the input of the neural network

self.net.setInput(blob)

# Get the output layers of the network

layers = self.net.getUnconnectedOutLayersNames()

# Feed the data to the network and get the output

outs = self.net.forward(layers)

# Call the postProcess() function from the Processing class

objects = self.postProcess(frame, outs)

# Draw the counting texts in the frame

for obj in objects:

label = obj [0]

confidence = obj

xmin = obj [2]

ymin = obj [3]

xmax = obj [4]

ymax = obj [5]

cv2.rectangle(frame, (xmin, ymin), (xmax, ymax), (0, 255, 0), 2)

cv2.putText(“Kamionete, Autobus, Kamion, Veture, Biciklete, kem-besore “, frame, label + “ “ + str(round(confidence, 2)), (xmin, ymin–5), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0, 255, 0), 2)

# Write the counting data into a csv file

self.writeData(objects)

# Show the frame

cv2.imshow(“Real Time”, frame)

# Press ‘q’ to quit

if cv2.waitKey(1) & 0xFF == ord(‘q’):

break

# Release the video capture device and destroy all windows

cap.release()

cv2.destroyAllWindows()

Measurement data from recordings through cameras and the application, for several 15 min intervals, for different visibility conditions, for one intersection, are presented in Figure 4, Figure 5 and Figure 6 as well as Table 1, Table 2 and Table 3.

Figure 4. Traffic count from camera and data processing in program (05:00–05:15).

Figure 5. Traffic count from camera and data processing in program (time: 19:00–19:15).

Figure 6. Traffic count from camera and data processing in program (time: 21:00–21:15).

Table 1. Results obtained from camera count in the morning.

Direction	Car	Motorcycle	Bus	Trucks	Total
Time period: 05:00–05:15
Up	19	0	0	2	21
Down	14	0	2	3	19
Total	33	0	2	5	40

Table 2. Results obtained from camera count during the day.

Direction	Car	Motorcycle	Bus	Trucks	Total
Time period: 19:00–19:15
Up	149	0	0	9	158
Down	224	1	7	22	254
Total	373	1	7	31	412

Table 3. Results obtained from camera count in the evening.

Direction	Car	Motorcycle	Bus	Trucks	Total
Time Period: 21:00–21:15
Up	53	1	0	3	57
Down	125	5	2	23
Total	178	6	2	26	212

The results of automatic measurements from the video recording for the period 05:00–05:15, for different categories of vehicles in traffic are presented in Table 1.

The results of automatic measurements from the video recording for the period 19:00–19:15 (Figure 5), for different categories of vehicles in traffic are presented in Table 2.

The results of automatic measurements from video recording (Figure 7) for the period 21:00–21:25, for different categories of vehicles in traffic are presented in Table 3.

Figure 7. Road traffic network in “Synchro” model.

The summary results, as shown in Table 4 and Table 5 and Figure 8, Figure 9 and Figure 10, prove the quality of the data obtained from the automatic count in terms of accuracy of measurements and adequate categorization of vehicles participating in the traffic.

Table 4. Results obtained from camera count in the different time periods.

Direction	Car	Motorcycle	Bus	Trucks	Average Difference in Counting	Processing Time of the Data
Time Period: 21:10–21:25
Manual	195	7	0	30	9.43%	15 min
Automatic	178	6	2	26
Time Period: 19:00–19:15
Manual	427	1	9	36	4.87%	15 min
Automatic	412	1	7	31
Time Period: 05:00–05:15
Manual	40	0	3	3	4.54%	15 min
Automatic	38	0	2	4

Table 5. Result obtained by “Synchro” model.

Movement	EBL	EBT	WBT	WBR	SBL	SBR
Lane configurations
Traffic volume (vph)	20	178	195	110	98	100
Future volume (vph)	20	178	195	110	98	100
Ideal flow (vphpl)	1900	1900	1900	1900	1900	1900
Total lost time (s)	4.5	4.5	4.5	4.5	4.5
Lane util, factor	1.00	0.95	0.95	1.00	1.00
Frpb, ped/bikes	1.00	1.00	1.00	0.45	0.83
Flpb, ped/bikes	1.00	1.00	1.00	1.00	1.00
Frt	1.00	1.00	1.00	0.85	0.93
Flt protected	0.95	1.00	1.00	1.00	0.98
Satd, flow (prot)	1770	3539	3539	714	1406
Flt permitted	0.95	1.00	1.00	1.00	0.98
Satd, flow (perm)	1770	3539	3539	714	1406
Peak-hour factor, PHF	0.92	0.92	0.92	0.92	0.92	0.92
Adj. flow (vph)	22	193	212	120	107	109
RTOR reduction (vph)	0	0	0	48	29	0
Lane group flow (vph)	22	193	212	72	187	0
Confl. peds. (#/hr)	220			200	200	200
Turn type	Prot	NA	NA	Perm	Prot
Protected phases	7	4	8		1
Permitted phases				8
Actuated green, G (s)	20.5	100.5	75.5	75.5	16.5
Effective freen, g(s)	20.5	100.5	75.5	75.5	16.5
Actuated g/C ratio	0.16	0.80	0.60	0.60	0.13
Clearance time (s)	4.50	4.50	4.50	4.50	4.50
Lane grp cap (vph)	287	2822	2120	427	184
v/s ratio prot	c0.01	0.05	0.06		c0.13
v/s ratio perm				c.0.10
v/c ratio	0.08	0.07	0.1	0.17	1.02
Uniform delay, d1	44.7	2.7	10.8	11.3	54.8
Progression factor	1.00	1.00	0.87	0.58	1.00
Incremental delay, d2	0.5	0	0.1	0.80	71.1
Approach delay (s)		7.1	8.7		125.9
Approach LOS		A	A		F
Intersection Summary
HCM 2000 control delay	41.4		HCM 2000 level of service			D
HCM 2000 volume to capacity ratio	0.28
Actuated cycle length (s)	126		Sum of lost time (s)			13.5
Intersection capacity utilization	39.60%		ICU level of service			A
Analysis period (min)	15
Critical lane group

With colored numbers are selected as the main indicators of traffic.

Figure 8. Difference between automatic and manual vehicle counts (time 21:10–21:25).

Figure 9. Difference between automatic and manual vehicle counts (time 19:00–19:15).

Figure 10. Difference between automatic and manual counts of vehicles (time 05:00–05:15).

Referring to the results obtained from the automatic method and manual counts performed by personnel employed to count vehicles, the largest changes in data quality are observed in the period of limited overnight visibility (time 21:10–21:25).

In this period, the counting data had a deviation of about 9.5%, while in other measurement periods, when visibility was good, the differences between the manual counting method and the automatic method through the applied program followed a deviation of about 4.87% for the period 19:00−19:15 and 4.54% for the period of measurements 05:00–05:15. In the quality of automatic measurements and accurate categorization of vehicles, the position of the cameras is also important; in this case, they were put at a certain angle to the direction of movement of the vehicles; therefore, the accuracy loss was likely to be greater. These deviations are more pronounced especially in the process of categorizing vehicles, due to the reduction in opportunities for accurate recognition and measurement of the dimensions of these vehicles.

3. Impact on Qualitative Traffic Indicators

The analysis and evaluation of the qualitative indicators of traffic operation are based on the basic data obtained in the research, traffic surveys or recordings; therefore, it is very important that this data be of the highest quality. In the concrete case of this research, based on the maximum difference of 9.5% between the sets of data obtained from manual and automatic counts, for the period with limited visibility at night (21:10–21:25), we used the “Synchro” model to analyze the impact of this on the results of important indicators of traffic operation such as: lane utilization factor, standard flow, adjusted flow, lane group flow, v/s ratio, v/c ratio, progression factor, approach delay and approach LOS. The construction of a road network in the model “Synchro”, including traffic modes and certain traffic demands by means of manual and automatic methods of measurements as well as the simulation of traffic operation, is presented in Figure 7 and Figure 11.

Figure 11. Road traffic simulation in “Sim” model.

The results of the qualitative traffic indicators obtained from the "Synchro" model are shown in Figure 8. Referring to the results obtained, no significant changes were observed in the results of traffic indicators which were analyzed through the "Synchro" model (Figure 8); therefore, we can conclude that both the quality, and the accuracy of the data obtained from the use of the automatic method through the cameras and the application of the analytical algorithm formulated in the programming language "Python" are high. As such, we recommend this model for use for the purpose of traffic counts and categorization of vehicles circulating in traffic, because even in the most extreme conditions with the possibility of greater accuracy loss in results at night (conditions of limited visibility), the impact of these losses on the analysis of important traffic indicators is negligible.

4. Conclusions

The analysis and evaluation of the qualitative indicators of traffic operation are based on the basic data obtained in the research, traffic surveys or recordings; therefore, it is very important that this data be of the highest quality. Referring to the results obtained from the automatic method (Python algorithm application) and manual counts performed by personnel employed to count vehicles, the largest changes in data quality were observed in the period of limited overnight visibility (time 21:10–21:25). In this period, the counting data had a deviation of about 9.5%, while in other measurement periods, when visibility was good, the differences between the manual counting method and the automatic method through the applied program followed a deviation of about 4.87% for the period 19:00–19:15, 4.54% for the time period of measurements 05:00–05:15. These deviations are more pronounced especially in the process of categorizing vehicles, due to the reduction in opportunities for the accurate recognition and measurement of the dimensions of these vehicles. Referring to the results obtained, no significant changes were observed in the results of the traffic indicators which were analyzed through the “Synchro” model; therefore, we can conclude that both the quality, and the accuracy of the data obtained from the use of the automatic method through the cameras and the application of the analytical algorithm formulated in the programming language “Python” is high. This model provides accurate results even during traffic counts in conditions with poor visibility at night. As such, we recommend this model for use for the purpose of traffic counts and the categorization of vehicles circulating in traffic because, even the most extreme conditions giving the possibility of a greater accuracy loss in results at night (conditions of limited visibility), the impact on the analysis of important traffic indicators of these losses is negligible. The authors′ next study will focus on developing a methodology for measuring the accuracy of traffic counts and traffic quality indicators in adverse weather conditions such as snow, rain, and fog using edge computing and IoT.