Autonomous Navigation Technology for Low-Speed Small Unmanned Vehicle: An Overview

Abstract

In special locations (scenes) such as campuses and closed parks, small unmanned vehicles have gained more attention and application. Autonomous navigation is one of the key technologies of low-speed small unmanned vehicles. It has become a research hotspot, but there are still many problems, such as perception sensitivity, navigation, and positioning accuracy, motion planning accuracy, and tracking control accuracy. In order to sort out the research status of the key technologies of autonomous navigation for small unmanned vehicles more clearly, this paper firstly reviews the key technologies of autonomous navigation and presents an analysis and summary. Finally, future research trends of small unmanned vehicles with low speed are given.

1. Introduction

An unmanned vehicle is a kind of vehicle that can sense the environment and navigate autonomously without human operation. It uses millimeter wave radar, lidar, GPS, and a camera to detect the surrounding environment, understands the sensing information, identifies obstacles and traffic signs through the scene perception system, and makes appropriate path planning decisions [1]. Unmanned vehicles break through the traditional driving mode, which can effectively improve the driving safety and stability of vehicles and reduce the incidence of traffic accidents. It is an important way of intelligent travel in the future. At present, with the development of unmanned technology, unmanned vehicles have been gradually applied in cities [2], agriculture [3], industry [4,5], special operation environments [6], and other fields. Compared with the complexity of cities, there are a lot of demands in closed environments (such as airports), which are easy to be applied in such environments [7].

Low-speed self-driving vehicles gradually mature with the progress of driverless vehicles. Since 2010, the demand for low-speed driverless vehicles in closed environments, such as sanitation, logistics, ports, and other fields, has increased rapidly. Major countries around the world have also actively deployed in the field of low-speed driverless vehicles, and vigorously promoted the application solutions in specific scenarios. Among them, the application of low-speed small driverless vehicles is becoming more and more common. Specifically, low-speed small driverless vehicles are different from those used in urban environments. The driving environment is relatively closed and the speed is less than 45 km/h, especially in low-speed application scenarios such as airports, ports, mining areas, sanitation, terminal distribution, campuses, scenic spots, and closed parks; it is divided into three categories: manned, cargo and special vehicles. As shown in

Table 1. Types of low-speed small unmanned vehicles.

Category	Species
Carry human beings	Campus bus [8]
	Scenic spot sightseeing bus
	Park tour bus
Cargo type	Express delivery vehicle [9]
	Workshop material transfer truck [10]
	Wharf transport vehicle [11]
Special purpose vehicle	Sanitation vehicles (sweepers [12], snow removal vehicles [8], high-speed marking vehicles, etc.)
	Patrol car (community, airport, etc.)

.

(1) Low-speed small unmanned vehicles carrying human beings

As the driverless cars of General Motors, Waymo, Uber, Google, and Tesla gradually matured, driverless technology turned its attention to public scenes such as school campuses, residential communities, and office parks. After 2007, a number of driverless low-speed bus vehicles represented by EasyMile, May Mobility, and Navya appeared one after another, and then unmanned bus manned services were launched in China, Japan, the United States, Germany, and Australia. The most likely use for unmanned manned buses is to supplement or expand the public transportation system. It is especially suitable for meeting the travel needs of major areas such as university campuses, retirement communities, and entertainment or business districts [7]. Navya’s unmanned low-speed bus is 4.7 m long and 2.1 m wide, can carry 15 passengers, has a maximum speed of 45 km/h, and has been tested at a speed of 25 km/h on public roads [7]. On 1 September 2016, the world’s first unmanned bus line was put into trial operation in Dubai. This small electric driverless bus is called EZ10, so it is feasible for driverless buses to come earlier than smart cars [13].

(2) Low-speed small unmanned vehicles carrying cargo

Logistics distribution field: In the development process of the modern logistics industry, unmanned distribution logistics service based on unmanned driving technology [14] has become the development trend [15]. Although there are still many limitations in the current development process, unmanned distribution logistics services will become the main form of logistics industry operations in the future. From the global application situation, unmanned distribution is the most widely used field of low-speed automatic driving. Starship Technologies, a British company, made the unmanned delivery vehicle earlier. Its unmanned delivery vehicle is equipped with nine cameras and a complete obstacle avoidance system, which can completely perform tasks automatically, drive at a speed of four miles per hour, and can transport 20 pounds (about 9 kg) of goods at a time. Unmanned distribution is the most widely developed in the United States at present. The Nuro Company of the United States also launched the low-speed unmanned distribution vehicle R1 in December 2018. In addition, Robby Robot of RobbyTechnologies in the United States and CarriRo Delivery released by ZMP, a Japanese robot company, have also conducted research on low-speed small unmanned delivery vehicles. The fifth generation intelligent express vehicle was released by China Jingdong Logistics and the unmanned delivery vehicle with L4 level was launched by Meituan. With the outbreak of the COVID-19 epidemic, the demand for “unmanned economy” scenes has been continuously expanded, and unmanned vehicles can replace manual labor to complete the “last mile” distribution work to a certain extent. Many enterprises are using unmanned vehicles to distribute materials and support the anti-epidemic.

(3) Special low-speed small unmanned vehicles

Sanitation cleaning field: With the development of artificial intelligence, an unmanned sweeper is proposed, which can realize automatic driving cleaning and overcome obstacles independently [16]. For large-scale cleaning operations and cleaning work in severe rain and snow, unmanned driving can not only improve the working environment of sanitation personnel, and improve work efficiency, but also greatly reduce the operating cost of sanitation cleaning. In addition, indoor cleaning robots have been developed earlier. After the development of unmanned technology, unmanned small sweepers in outdoor scenes have also been first applied. For example, Envay, a German driverless company, and FYBOTS, a French company, have successively withdrawn from driverless sanitation vehicles, while well-known enterprises in China include Wo Xiaobai, an unmanned sweeper released by Smart Walker.

Security field [17]: For airports, closed parks, campuses, and communities, low-speed small unmanned vehicles can effectively replace patrol personnel, realize 24-h patrol inspection, and effectively ensure the safety of public places. For example, the unmanned patrol car launched by Singapore Otsaw Company and the unmanned patrol car launched by UISEE were applied to Hong Kong International Airport in November 2021.

To sum up, due to its unique system structure, a low-speed unmanned vehicle can not only realize conventional driving behavior but also realize special driving abilities such as autonomous environment perception, motion planning, and navigation control. It can effectively improve the safety and reliability of vehicles, reduce the fatigue and labor intensity of drivers, and make low-speed unmanned vehicles such as small sweepers, small scenic sightseeing buses, and small logistics distribution vehicles become the research hotspot of researchers. Its development also has broad application prospects and scientific research value. Unmanned vehicles integrate image processing, automatic control, artificial intelligence, and many other technologies to realize the autonomous navigation of vehicles. In [18], a neural network model to solve the problem of autonomous navigation on the ground is developed to improve the ability of navigation in complex terrain environments. As to whether unmanned vehicles can be utilized in a complex environment, autonomous navigation is the core evaluation standard to mark their intelligence degree.

The so-called autonomous navigation means that unmanned vehicles can sense the external environment through their own sensors, can plan their routes autonomously, and avoid dynamic and static obstacles when encountering dynamic and static obstacles [19]. Therefore, autonomous navigation mainly solves three problems here: First, where am I? The second is where I go, and the third is how I should go. Finally, the unmanned vehicle can reach its destination autonomously and safely. If we want to solve these three problems, we need to study the autonomous navigation technology of unmanned vehicles. Generally speaking, the current mainstream autonomous navigation schemes are divided into visual navigation [20,21] and laser navigation. In [22], a new PID autonomous navigation system for intelligent vehicles based on ROS and lidar is proposed, which uses lidar to scan the environment and build a map, and realizes visualization through Rviz. The PID algorithm is integrated into ROS, and the deviation correction accuracy is improved by filtering. It has a good effect on map construction and autonomous navigation and is superior to traditional amcl navigation in autonomy and efficiency. In [23], a light autonomous vehicle system VOLWA based on pure vision is proposed, which adopts the visual position recognition algorithm of removing dynamic targets by channel selection, effectively improves the performance of position recognition, and the autonomous navigation system runs well. In addition to the above two independent sensor autonomous navigation systems, there are also multi-sensor fusion methods such as lidar and vision fusion methods to realize autonomous navigation systems [24,25].

Generally speaking, autonomous navigation technology mainly includes environment awareness technology, map building technology, motion planning technology, and tracking control technology. At present, there are few overview papers on autonomous navigation technology of low-speed small unmanned road vehicles. This paper mainly reviews and analyzes the key technologies of autonomous navigation of road vehicles, and puts forward the prospect of future development.

The rest of this article is as follows: Section 2 reviews environment awareness technology, Section 3 reviews map building technology, Section 4 reviews navigation and positioning technology, Section 5 reviews motion planning technology, Section 6 reviews tracking control technology, and Section 6 summarizes this article.

2. Review of Environmental Perception

At present, most teams engaged in the research of driverless vehicle technology focus on perception or control, and the research focus is mainly on the driverless field at medium and high speed, while there are relatively few driverless researchers at low speed and in a specific environment. In order to achieve the goal of autonomous navigation, low-speed small unmanned vehicles first need to be able to understand the state information of the vehicle itself and the environmental information around the vehicle. The acquisition of this information requires the vehicle sensing system to provide the basis for system decision-making, which is also one of the key technologies of autonomous navigation systems [26]. The environmental sensing system includes internal state acquisition sensors and external sensors for environmental sensing [27].

The state information of unmanned vehicles mainly includes the speed, acceleration, inclination angle, position, and other information of the vehicles [28]. This kind of information is mainly measured by inclination sensors, gyroscopes, and other sensors. In

Table 2. Sensor characteristics of vehicle itself.

Sensor	Function	Characteristic	Precision	Disadvantages
GPS	Navigation and positioning sensors	Geolocation and time information	Centimeter scale	signals are blocked by high-rise buildings, and so on.
IMU	Control and Navigation	force, violent rate and magnetic field.	Centimeter scale	Affected by accumulated errors, resulting in drift
Encoders	An analog or digital signal	position, direction and velocity	Meter level	Errors often occur in measurement

, the vehicle’s own attitude sensor characteristics are shown [29].

The external environment sensing system requires external sensing sensors, including lidar, radar, and ultrasonic sensors, as well as monocular, stereo, omni-directional, infrared, and event cameras [30].

In

Table 3. Vehicle External Sensor Characteristics.

Sensor	Detection Distance (m)	Accuracy	Function	Advantages	Disadvantages	Unmanned Applicable Vehicle Type
						Bus	Delivery Vehicle	Sanitation Vehicle	Patrol Car
Millimeter wave radar [31]	<250	Medium	detect the position and speed of the target	No environment affected, detection distance long	Small detection angle	√	√		√
Lidar	<200	High	detect the position and speed of the target	Long detection distance, wide field of view, high data acquisition accuracy, no lighting conditions affect	Bad weather, the performance will decline			√
Ultrasonic	<5	Low	target detection	Low cost and small volume	Low precision, narrow visual field and blind spot			√
Monocular Camera	-	High	target detection	Image has color, texture and high resolution	Affected by weather and lighting conditions	√	√	√	√
Stereo Camera	<100	High	Distance estimation	Get color and motion information	Vulnerable to weather and lighting conditions, narrow vision	√	√	√	√
Omni Direction Camera	-	High	Slam and 3D reconstruction	Great vision	Vulnerable to weather and lighting conditions, High computing power	√
Infrared Camera	-	Low	object detection	Good performance at night	No color or texture information, low accuracy	√			√
Event Camera	-	Low	object detection	Dynamic imaging	Affected by weather and lighting conditions	√

, external sensors are briefly shown [30].

To sum up, low-speed driverless vehicles need to choose different sensor configuration schemes according to different application scenarios. Although most sensor configurations are still the same as those of driverless vehicles used on expressways, considering the particularity of low-speed driverless vehicles, there are some differences between onboard sensors and high-speed driverless vehicles.

Low-speed small unmanned vehicles carrying human beings: In the configuration of sensors, sensors such as IMU, GNSS, lidar, encoder, and camera are routinely selected. However, considering that the environment in which vehicles run does not have the complexity of urban roads, and does not exist in densely populated areas and long-distance highways, the cost will not be very high. The selection of sensor configuration parameters reflects the advantages of economy, and the performance and quantity of sensor configuration are obviously different from those of high-speed unmanned vehicles [32].

Low-speed small unmanned vehicles carrying cargo: In the configuration of sensors, considering the use scenarios, the on-board sensor configuration of driverless vehicles carrying goods also reflects the advantages of economy, and considering the high probability of not working at night, the implementation complexity is smaller; however, considering the human-computer interaction function between cargo unmanned vehicles and recipients, it is necessary to add additional sensors to the vehicle-mounted electronic control system.

Special low-speed small unmanned vehicles: Considering also that unmanned sweepers need to collect garbage and unmanned inspection vehicles need to detect and identify targets, the sensing system scheme based on visual sensors can be considered in sensor configuration, and expensive sensors such as lidar can not be considered in campus and closed park environments. For example, an infrared camera can be used. As shown in Table 3, the infrared camera has some disadvantages. To solve the problem of low resolution, in [33], it is mentioned that enhancing the contrast of infrared images, adopting the yolov3 neural network, and building a stereoscopic infrared vision system are effective methods to solve the detection problem. In [34], it is mentioned that the fusion of visible light camera and infrared camera can be used to obtain information-rich images, and the infrared image and laser radar point cloud can be used to obtain the depth and position information of the target, These initiatives may overcome the shortcomings of infrared cameras.

The environmental perception of unmanned vehicles mainly includes: (1) road state information, and (2) pedestrian, vehicle, and other obstacles detection.

(1) Road state information perception

Mainly for the identification of the road environment of low-speed unmanned vehicles. Because the current urban road environment is complex, for unmanned vehicles, the environment is complex, the requirements are high, and the potential safety hazards are quite numerous. Expensive equipment leads to high costs. Therefore, for low-speed driverless vehicles, different road states such as closed or semi-closed park roads, campus roads, scenic roads, and community roads are mainly identified. Different roads may have complex road pavement environments such as road smoothness and road water accumulation; because of the characteristics of low-speed unmanned vehicles, there are few lane lines, ground traffic signs, and traffic lights. The perception of road state information mainly depends on the camera and adopts the way of visual recognition.

At present, the perception of vehicle road state in urban scenes mainly focuses on lane detection, traffic signs, and traffic signals, while the research on road surface state is less, and the related data sets are mainly for the data of test sections or urban roads, while the data sets for complex road conditions are less.

Specifically, for the low-speed small unmanned vehicles studied in this paper, according to different application scenarios, different vehicles have different perception requirements for road state information. The details are as follows:

Low-speed small unmanned vehicles carrying human beings: These kinds of vehiclesare mainly used as campus buses or scenic sightseeing buses on campuses or in closed parks. Different from unmanned vehicles used on conventional urban roads, they mainly face unstructured or semi-structured roads, so it is necessary to consider identifying and analyzing road infrastructure conditions such as road conditions [35], for facing the rugged road surface, it is necessary to consider configuring high-efficiency energy regenerative shock absorbers [36].

Low-speed small unmanned vehicles carrying cargo: For small unmanned vehicles in the logistics field, because the working environment is mostly unstructured or semi-structured roads such as communities, it is important to consider identifying the road status of communities to avoid vehicle failure caused by potholes and protrusions [37].

Special low-speed small unmanned vehicles: For unmanned vehicles in the field of sanitation, because most of the use scenarios are unstructured or semi-structured road conditions, it is necessary to consider the potholes, protrusions, accumulated water, and ice on the road surface in the process of garbage cleaning, so as to avoid incomplete garbage collection or wrong garbage collection identification (such as invalid garbage such as accumulated water and glass bottles) during the operation of the sweeper. For unmanned inspection vehicles in the field of security, it is necessary to focus on the road environment such as potholes, protrusions, accumulated water, and ice on the road surface, so as to avoid skidding of inspection vehicles and vehicle failure.

(2) Target detection around pedestrians, vehicles, and other vehicles

Due to the characteristics of low-speed unmanned vehicles’ application scenarios, target detection tasks include vehicle [38], pedestrian, non-motor vehicle, and general obstacle detection.

There are two main means of environmental perception: visual perception and radar perception [39]. For low-speed unmanned vehicles, visual perception has become a widely used environmental perception method because of its low input cost, simple operation, and large amount of information. Image segmentation and object detection are two key technologies in visual perception systems [40]. The main form of image segmentation is semantic segmentation [41]. In visual perception technology, a semantic segmentation algorithm is usually used to segment background categories such as sidewalks, sky, and buildings because of its richer image understanding, so as to extract the workable area of unmanned vehicles. Target detection technology is widely used to detect pedestrians, vehicles, and other targets in the perceptual picture because of its higher detection speed and the characteristics of obtaining the position and number of perceptual objects in the scene. Recently, Deep Neural Network (DNN) has been used to solve the problem of object detection and classification, road detection, and image semantic segmentation in automatic driving perception [42,43,44]. At present, the frequently used classic networks include Fast R-CNN, R-FCN, mask RCNN, yolov5, SSD, etc., all of which have achieved good results on Kitti and other automatic driving road data sets.

In [45], a Yolov5 neural network model is adopted, which has the characteristics of fast training speed, easy installation and use, strong robustness, and high recognition accuracy under severe weather conditions, which is beneficial to improving the accuracy of unmanned vehicle target detection. Another commonly used algorithm model is SSD, which is used in [46] which has fast detection speed and good real-time performance, but the model is difficult to converge, which has certain limitations on small target detection effects.

Radar sensing mainly includes active ranging sensors such as millimeter wave radar and lidar radar, which are realized by multi-sensor fusion. Because the fusion of lidar, millimeter wave radar, ultrasonic, and other sensors can better meet the sensing needs of low-speed unmanned vehicles under complex roads and bad weather conditions, compared with the single laser point cloud data, the data processing capacity can be reduced and the real-time performance can be improved through the fusion processing algorithm. At the same time, whether there are obstacles or not can be obtained through the fusion data during trajectory planning, without knowing the specific information of obstacles. For lidar, it collects the point cloud information of the environment, which contains accurate three-dimensional coordinates and reflectivity of the object surface, which is of great help to determine the accurate position of the target in three-dimensional space. However, the point cloud data itself is sparse and disordered, which makes it more difficult to process and store the point cloud compared with the image. Moreover, the point cloud lacks color information, which makes it difficult to judge the category of the target and identify the dynamic target with color.

At present, most visual inspection algorithms are based on a deep convolution neural network, and its internal principle is convolution operation, and more convolution operations may have the problem of insufficient computing power; in fact, most visual data in the field of autonomous driving is a useless background in processing. Traditional visual processing algorithms still need to consider the whole background, even if most of the background areas have not changed, this processing algorithm wastes a lot of computing power and time. Event cameras have overwhelming advantages in target tracking, motion recognition, and other fields, especially suitable for automatic driving. The emergence of an event camera has successfully solved the problem of visual processing in automatic driving.

Visual perception is an important means to perceive the external environment of low-speed small unmanned vehicles. Facing complex weather conditions, it is difficult to realize complex weather environment perception only by a single sensor. Other sensors can be integrated through visual sensors. The typical scene is to integrate sensors such as radar and camera on foggy and snowy days [47], as this overcomes the shortcomings of a single sensor in complex weather, and can solve the problem of complex weather environment perception to a certain extent. Vision sensor plays an irreplaceable role in road target recognition and scene understanding, and vision is also an essential means in road damage and water accumulation in the complex road environment. Therefore, the fusion of vision and lidar point cloud information has higher detection accuracy and good real-time performance than a single recognition method in obstacle recognition [48,49], and various sensors often need to work together and complement each other’s advantages to create an environment perception solution for automatic driving [50,51]. In

Table 4. Sensor Fusion Technologies.

Sensor Combination	Realization Function	Characteristic
Vision-LiDAR/Radar	Localization, object detection and environment modeling	High calculation efficiency and modeling accuracy
Vision-LiDAR	Dynamic object Tracking	Avoid trajectory loss caused by LIDAR
GPS-IMU	Absolute localization system	Cumulative error is reduced and the calculation amount is small
Vision-Odometry	Localization	Frequency: 35 Hz Average error: 34 cmAngle error: 1–3 degrees.
Odometry-Magnetic Sensor	Localization	Better results in attitude estimation.
Stereo vision- IMU [53]	Localization	Improve the accuracy of state estimation
GPS, Odometry, Inertial, Laser Sensor [54]	Location estimation	Small estimation error
LIDAR, IMU, Wheel Odometry [55]	Localization	Vehicle moves rapidly, error will increase
Radar and Infrared [56]	Multi-object tracking	Phase delay of target measurement data

, the main Sensor Fusion Technologies are briefly shown [30,52].

To sum up, it can be seen that the target detection technology around low-speed small unmanned vehicles is the same as that of ordinary unmanned vehicles such as urban high-speed. The difference may be that the types of obstacles to identifying targets are slightly different, and the focus of detection technology may be different to some extent, which is also determined by the simple application scenarios of low-speed unmanned vehicles.

Low-speed small unmanned vehicles carrying human beings: The application scenarios are mainly in non-urban environments such as campuses and parks, with few traffic lights on the route, driving along a fixed route, only running during the day, stable environment and no complex urban buildings [57], mainly based on visual perception, supplemented by radar perception. In special weather environments such as rain, snow and fog, the scheme of visual plus radar fusion perception can be adopted to ensure the safety and reliability of vehicle operation.

Low-speed small unmanned vehicles carrying cargo: For unmanned distribution vehicles, it is necessary to consider the complexity of traffic conditions in the driving route. At present, Zhongtong unmanned distribution vehicles developed by China Matrix Data Technology adopt four 32-line lidar schemes, which realize long-distance ranging and high-precision environment perception. The driving scene of the unmanned delivery vehicle in JD.COM, China is a non-motor vehicle lane with a speed of 15 km/h. The radar and image pre-fusion algorithm PAI3D is adopted to realize 3D target detection, which overcomes the problem of inaccurate depth estimation of 3D detection only by monocular vision, and reduces the problems of false detection and missed detection of obstacles [58]. China SF unmanned vehicles are applied on campus, carry out contactless distribution services, adopt the scheme of vision and radar fusion perception, have the ability of 120 m distance perception, and can operate normally under complex weather conditions such as foggy days, rainy days and nights.

Special low-speed small unmanned vehicles: In the field of sanitation, for unmanned sweepers, it is necessary to consider the identification and detection of road garbage classification, with emphasis on early warning of garbage that cannot be collected to avoid equipment failure, and in [59], two Mask R-CNN neural network models are constructed to realize automatic identification of road garbage. In [60], two algorithms, namely, the garbage type identification algorithm and road garbage coverage algorithm based on the Faster-RCNN model, are proposed to identify road garbage and realize high-efficiency and energy-saving automatic cleaning. In [61], an unmanned sweeper adopts a centralized multi-sensor fusion algorithm based on vision and radar sensors to detect pedestrians, vehicles, and other multi-targets. In the field of security, unmanned patrol cars need to focus on pedestrians and vehicles, and [62] uses thermal images, fast contour models and particle filters to detect, track and identify pedestrians in real-time.

For the construction of an environment awareness system for low-speed small unmanned vehicles, considering that it is still in the stage of demonstration application, this is reflected in the importance of low-speed small unmanned vehicles during the pneumonia epidemic in COVID-19, and the price factor is not the main problem of scene application. In the future, with the popularization of large-scale applications, the cost of vehicle-mounted lidar, millimeter wave radar, and the camera will gradually decrease. Sensors usually applied to high-speed unmanned vehicles on urban roads will be applied to low-speed small unmanned vehicles, which will also improve the safety and reliability of the sensing system of low-speed small unmanned vehicles.

3. Review of Map Building

Mapping is the description of the environment in the process of vehicle movement, and the forms of mapping are divided into metric maps and topological maps. A metric map accurately represents the positional relationship of objects on the map; divided into sparse maps and dense maps.

Among them, sparse maps express the environment abstractly and cannot express all the information of the surrounding environment, while dense maps are usually composed of many small blocks according to a certain resolution, but they consume a lot of storage space and may fail. Topological maps are composed of nodes and edges, which mainly express the connectivity between map elements, and are not good at expressing maps with complex structures [63]. In recent years, the mapping method based on multi-sensor fusion, such as lidar, camera, and IMU fusion, has gradually become a focus of development. Through the information of various sensors, the robustness of mapping of low-speed small unmanned vehicles can be increased, and more accurate and informative maps can be obtained. In addition, due to the serious positioning errors caused by GPS unlocking, SLAM technology can complete map construction for unmanned vehicles in an unknown environment, which is the basis of autonomous navigation.

(1) Multi-sensor fusion technology to build high-precision maps. Lidar is used as the main sensor, dot matrix cloud scanning, and then the images collected by the auxiliary camera are fused together, and then the data are analyzed, such as traffic signs, street signs, etc., and finally, the scene map is synthesized. Based on multi-sensor fusion technology, outdoor high-precision map scenes can be constructed. It can be applied to typical low-speed application scenarios such as factories, communities, and schools.

(2) SLAM map construction method. As the core of autonomous navigation for low-speed small unmanned vehicles, high-precision maps can gradually build maps with global consistency in an unknown environment without prior information. At present, the sensors commonly used in SLAM map construction are lidar and visual cameras, which correspond to the laser SLAM algorithm and visual SLAM algorithm, respectively.

Visual SLAM can be divided into monocular, binocular, and RGBD types according to different cameras. The biggest problem with a monocular camera is that it cannot obtain accurate depth information, and it is difficult to calibrate the binocular camera. It consumes computing power to calculate depth information by using pixel values. The RGBD camera can obtain environmental information through its own laser transmitter without consuming a lot of computing power to calculate depth information, but its small measurement range and large noise lead to an inability to apply to large-scale SLAM of unmanned vehicles, and SLAM realized by vision is easily affected by illumination; Laser SLAM has high measurement accuracy, strong anti-interference ability, wide measurement range, fast ranging speed and is not affected by illumination information, which is very suitable for obtaining scale information of the environment. However, the information obtained by lidar lacks texture and color information, and the realization of loop detection is relatively difficult. Therefore, it is necessary to complete the fusion of laser point cloud information and visual information and use the fused information to realize SLAM.

Mapping technology is one of the core technologies of unmanned vehicles. For low-speed unmanned vehicles, mapping technology may be different in different application scenarios.

Low-speed small unmanned vehicles carrying human beings: in [64] the medium and low-speed unmanned buses are mapped by laser SLAM, and the collected data are inherently robust to environmental conditions, which improves the accuracy of mapping.

Low-speed small unmanned vehicles carrying cargo: Version 4.0 of unmanned distribution vehicles in JD.COM, China uses the scheme of vision and radar integration and adopts IMU and SLAM technology to realize point cloud mapping and full scene mapping. Unmanned forklift adopts lidar 3D SLAM technology to realize real-time mapping of point cloud, which is simple to deploy and can be put into use quickly [65].

Special low-speed small unmanned vehicles: in [66], laser SLAM is used to scan and model the surrounding environment and generate point cloud images. However, for unmanned sweepers and patrol cars used in outdoor places such as campuses, communities, and parks, considering the economic requirements, it is necessary to focus on using visual SLAM to build maps in the future.

The application scenarios of low-speed small unmanned vehicles are mostly non-urban roads. The routes for such as unmanned sweepers, distribution vehicles and patrol cars are relatively fixed, and maps can be updated offline; in the future, the scheme of regional cloud server will be adopted, multi-vehicles will work together, and maps will be built and updated in real-time.

4. Review of Navigation and Positioning Technology

Low-speed small unmanned vehicles need to rely on their autonomous navigation system to achieve their own high-precision positioning, which is also the basis for vehicle motion planning and decision control.

Autonomous navigation and positioning technologies for low-speed small unmanned vehicles can be divided into three categories:

Relative navigation and positioning: Inertial navigation system (INS) and dead reckoning system (DRS) are commonly used systems for relative positioning, and their outputs are highly autonomous, almost free from external interference, and have good smoothness and short-term accuracy [67]. The main disadvantage of relative positioning is that the cumulative measurement errors will accumulate over time and tend to be unbounded. In [68], an algorithm for relative positioning and vehicle odometer using GPS carrier phase measurement is proposed. Single-frequency or dual-frequency GPS measurements are used to evaluate the accuracy of relative positioning algorithms. The accuracy of relative position vector estimation is centimeter level.

Absolute navigation and positioning: GPS measurement calculation is based on the pseudo-range sent by different satellites, which is one of the most widely used absolute positioning systems. It can provide three-dimensional position and velocity information without accumulated errors on a global scale, and obtain consistent accuracy over time [67]. However, GPS is easily occluded, which affects positioning accuracy, its output frequency is low, and its position data usually has random noise.

Integrated navigation positioning: Considering the advantages and disadvantages of relative positioning and absolute positioning, in order to make up for the shortcomings of the single positioning method. Usually, integrated navigation positioning schemes are adopted, such as GPS + IMU, GPS + odometer, and so on.

For low-speed unmanned vehicles, GPS navigation is suitable for mobile vehicles on outdoor open driving roads [69]. When buildings and trees are blocked, GPS signals are weak, resulting in an unlocked phenomenon, which cannot be used for positioning and navigation of unmanned vehicles. In order to realize the accurate positioning of vehicles in the unknown road environment, wheel odometer, visual odometer, laser odometer, and radar odometer and their combination of visual laser odometer and visual-inertial odometer are adopted [28]. With the development of computer technology in hardware and software, artificial intelligence (AI) related solutions have attracted more and more attention. When GPS works well, artificial intelligence will learn GPS/INS behavior patterns and build a model to map vehicle dynamics (attitude, speed, or position) and corresponding errors. Errors from the neural network model will be used to compensate for INS drift during GPS signal interruption, so as to meet the problem of positioning accuracy degradation caused by GPS lock loss [70].

To sum up, the application scenario of low-speed small unmanned vehicles is simpler than that of the urban environment, without too many high-rise buildings, and the problems of high-precision positioning are mostly caused by trees and other environmental occlusions.

Low-speed small unmanned vehicles carrying human beings: Navya’s unmanned buses rely on satellite-based positioning, while other manufacturers such as EasyMile use SLAM for positioning more [32].

Low-speed small unmanned vehicles carrying cargo: China Neolithic unmanned distribution vehicles adopt RTK positioning and lidar, and cooperate with the positioning scheme of high-precision maps. When there are trees and a weak GPS environment, they rely on lidar to cooperate with high-precision maps for positioning. However, China Smart Walker unmanned distribution vehicle adopts a differential GPS positioning system, and China Meituan unmanned distribution vehicle adopts multiple positioning technologies such as laser positioning, visual positioning, and semantic positioning to realize a multi-scene positioning system at the vehicle end. The unmanned delivery vehicle in JD.COM, China is based on high-precision stereo image data combined with the GNSS satellite positioning system to achieve high-precision positioning.

Special low-speed small unmanned vehicles: The unmanned sweeper developed by China Smart Walker adopts integrated navigation, laser, and visual SLAM fusion positioning scheme, which realizes high-precision positioning on unstructured roads. In [71], the unmanned sweeper uses laser point cloud, encoder data, and IMU data to locate the vehicle in combination, and locates the vehicle position through the measured data. In [72], the unmanned patrol car adopts a GNSS-RTK positioning system and receives GPS and BDS observation data at the same time. The number of available satellites is more than twice that of a single satellite system, so the positioning accuracy is significantly improved. In [73], a cost-effective GNSS/SINS integrated positioning system based on a fiber optic gyroscope (FOG) is designed. When satellite signals are blocked in a complex terrain environment, the test accuracy can meet the actual accuracy requirements of an integrated navigation system. When GNSS/RTK is used as the positioning system of an unmanned patrol car, the heading angle of GNSS may be affected due to the interference of the working scene. In [74], the electronic compass compensation method is adopted, and the sensor cost is reduced by more than 99%, which is of great significance for market application.

At present, one of the problems in the popularization and application of low-speed small unmanned vehicles is the price of a high-precision positioning scheme. Some of them are on campuses, scenic spots, communities, and other places. Low-speed small unmanned sweepers, distribution vehicles, patrol cars, and so on require higher economy, which cannot reach the higher price of urban road unmanned vehicles. With the large-scale application of high-precision GPS and other positioning sensors, the price will gradually decrease, and high-precision integrated navigation positioning sensors will be gradually configured to realize high-precision positioning on low-speed small unmanned vehicles.

5. Review of Motion Planning Techniques

Motion planning is to find a constrained path for moving vehicles between given two points. Motion planning includes path planning (space) and trajectory planning (time).

(1) Path planning

The path planning algorithm of low-speed unmanned vehicles needs to consider factors such as vehicle speed, road adhesion, vehicle steering angle, and external weather environments such as rain, snow, and fog, which is different from conventional indoor robots.

Path planning is the link between the perception system and decision control system of vulgar small unmanned vehicles, and it is the basis of autonomous navigation. The task of vehicle routing planning is to find a collision-free path from the initial state to the target state in a complex outdoor environment with obstacles.

Path planning can be divided into global path planning and local path planning. Global path planning is to determine the practicable and optimal path by using the known road information, such as obstacle position and road boundary, when the surrounding environment of the vehicle is known through the perception system and the map has been built. Under the guidance of the driveable area generated by global path planning, local path planning determines the trajectory of low-speed unmanned vehicles in front of the path according to the local environmental information perceived by sensors. Global path planning is mainly aimed at the scene where the vehicle environment is known, while local path planning is suitable for the scene where the vehicle environment is unknown.

Global path planning algorithms can be divided into four types: graph-based search methods, sampling-based methods, curve interpolation methods, and intelligent optimization methods.

① Method based on graph search

In this method, the state space is discretized into a search graph, and the optimal solution is calculated by various heuristic search algorithms. Dijkstra algorithm is one of the classical graph search algorithms. It can search the shortest path, but it will lead to too many unrelated nodes. Through the improvement of this algorithm, some more adaptive algorithms such as a * algorithm and LPA * algorithm are proposed. Its advantages are good real-time performance, fast global optimal (sub-optimal) path, and strong adaptability to dynamic and static environments in low dimensional space, but there are also problems of unreliable search performance in high dimensional space.

② Sampling-based path planning algorithm

Unlike the search-based approach, does not need to complete model the environment. In addition, this algorithm has the feature of random sampling, fast searching, and high efficiency planning. They have support for solving complex high-dimensional problems, but the computational efficiency is low. In [75], a dynamic avoidance obstacle control system of a small unmanned sweeper is completed by using the lattice programming algorithm, and an avoidance obstacle method that conforms to the driving characteristics of the sweeper is proposed.

③ Method of curve interpolation

This method fits the route curves of small unmanned vehicles with low speed under some safe and efficient conditions, such as polynomial curve, double circular arc curve, sine function curve, Bessel curve, B spline curve, and so on. Generally speaking, a polynomial algorithm mainly needs to consider geometric constraints to determine the parameters of curves. Based on parametric curves to describe the trajectory, this type of algorithm is more intuitive, but also can more accurately describe the road conditions that vehicles need to meet, and the planned trajectory is also very flat, curvature changes are continuous and can be constrained. Disadvantages include a large amount of computation, real-time performance is not very good, and its evaluation function is difficult to find the best parameter. Future research directions mainly focus on simplified algorithms and more perfect evaluation functions. At present, the curve fitting algorithm adopts a wide range of planning methods.

④ Swarm intelligence optimization algorithm

With the application of intelligent optimization technology, various swarm intelligence algorithms have been applied to path planning, which have the feature of self-learning, self-determination, and adaptive search, including GA, PSO, Firefly Algorithm, Artificial Fish Algorithm, ABC, and ACO algorithm.

Local path planning algorithm refers to real-time path planning to avoid obstacles when encountering obstacles, that is, autonomous obstacle avoidance algorithm. Typical local path planning algorithms include artificial potential field method and its variants, Bug algorithm, etc.

⑤ Algorithm based on deep reinforcement learning [76]

Without prior knowledge of environmental characteristics, low-speed small unmanned vehicles can use depth deterministic strategy gradient (DDPG) and reinforcement learning (RL) algorithm for continuous action space to navigate and exit an unknown environment full of obstacles [28,77]. That is to say, through the interaction and trial and error between onboard sensors and the surrounding environment, online knowledge learning is carried out, and the knowledge of the unknown environment around the vehicle is continuously acquired. In order to adapt to the environment and obtain the best decision-making action, the driving strategy of the vehicle’s dynamic path is improved, but on the other hand, it also has the problem of a high training cost.

In addition, for low-speed small unmanned vehicles, the current mainstream algorithms still have some problems, such as high marginal computing cost, weak constraint ability, and inability to measure the optimization degree of optimization objectives. In [78], a multi-objective programming model is proposed based on the demand of large-scale customer distribution, aiming at minimizing the overall resource consumption of distribution customers and optimizing customer satisfaction. The path planning of unmanned logistics distribution under an intelligent network is established.

(2) Trajectory planning

When the path planning process is to satisfy the lateral and longitudinal dynamics constraints of small unmanned vehicles at low speed, it becomes trajectory planning. The goal of trajectory planning is to plan a safe, reliable, and energy-saving trajectory for unmanned vehicles to complete the established driving tasks; to keep a proper distance from obstacles during driving to avoid a collision.

The typical is the optimal control method, it can obtain the path point’s time information [79]. In [80], a trajectory planning method is proposed, which considers boundary constraints, shortens the calculation time, and realizes fast trajectory planning. But, when have many obstacles in the vehicle driving environment, this method may be very time-consuming due to the need to solve complex optimization problems, and its inherent nonconvexity makes it difficult for the problem to converge without a proper initial guess.For low-speed unmanned vehicles, because the vehicle moving scene is not on a regular urban road, there may be various conditions such as uneven or accumulated water on the road surface, so for low-speed unmanned vehicles, it is necessary to consider the road surface state and make optimal motion planning.

Low-speed small unmanned vehicles carrying human beings: for low-speed unmanned buses in [81], the global path planning based on the Dijkstra algorithm is adopted to generate the optimal path trajectory, which can improve the accuracy of trajectory control, reduce the computational complexity of the controller and improve the driving efficiency of vehicles. For driverless buses, in [82], a framework successfully applied to autonomous electric buses is proposed. For global planning, ArcGIS is used to model the road network and plan the global optimal path. A cubic polynomial curve is used to generate a local trajectory, which reduces the calculation time. To sum up, the motion planning of low-speed unmanned vehicles carrying human beings needs to choose the best trajectory planning route for obstacles such as pedestrians and vehicles on the road, and safety and reliability are the first elements.

Low-speed small unmanned vehicles carrying cargo: for unmanned delivery vehicles, Ref. [83] adopts the strategy of combining ant colony optimization (ACO) with genetic algorithm (GA) to achieve the accuracy of vehicle routing planning. As for the economy of the last mile express delivery, Ref. [84] proposes a new route planning method based on clonal ant colony adaptive optimization (CAACO). The cost and time of fast delivery are lower than those of ACO, SA, and GA. To sum up, the factors of economy and reliability should be considered in the motion planning of unmanned distribution vehicles, and the surface conditions of roads such as communities and campuses should be comprehensively considered in the path planning, so as to avoid the problems of uneven potholes, snow-covered icy roads, vehicle skidding and other problems that easily lead to vehicle failures.

Special low-speed small unmanned vehicles: For the sanitation field, because the low-speed unmanned sweepers involve road cleaning operations, road conditions should be considered in path planning and trajectory planning, Ref. [85] proposes a road surface state function, which comprehensively considers various road conditions that sweepers may encounter, and realizes the optimal path that can avoid poor road conditions, instead of the shortest path that is normally considered. In the field of security, the shortest path needs to be considered more in the path planning of unmanned patrol cars, in [86], uses a fast exploration-based stochastic tree (RRT) for path planning, and the realization effect is close to the optimal solution controlled by numerical optimization based on nonlinear programming.

6. Review of Tracking Control Technology

Trajectory tracking control is one of the core technologies to realize the safe and stable operation of low-speed unmanned vehicles. According to the given vehicle path, the target path point is tracked in real-time, and the vehicle is controlled to move along the planned path to minimize the trajectory deviation. Through the research of PID control, sliding mode control, pure tracking algorithm, MPC model predictive control, robust control, and other [87,88,89,90,91,92] algorithms, the trajectory tracking accuracy can be improved and the expected target effect can be achieved.

Trajectory tracking is an important research direction for low-speed small unmanned vehicle control, and high-precision trajectory tracking control is particularly important for the safe driving of vehicles. Researchers have gradually devoted themselves to the research of trajectory tracking technology for low-speed unmanned vehicles. Because low-speed small unmanned vehicles work under complex road conditions such as campus and community, their working speed is much lower than that of existing urban traffic vehicles. Therefore, it is faced with how to reduce the tracking error and maintain the stability of the vehicle at different speeds under unpredictable working conditions and uncertain conditions. In the process of trajectory tracking control, however, the low-speed driverless vehicle is a complex nonlinear system [93], the coupling degree is high, so it is difficult to establish an accurate vehicle dynamics system model [94], and two-wheeled bicycles or four-wheeled robot models are often used instead of vehicles for research. Moreover, when the vehicle runs under different road conditions and climatic conditions, the parameters such as ground adhesion coefficient and road surface state are time-varying, which will affect the stable operation of low-speed unmanned vehicles. At the same time, low-speed unmanned vehicles will change with the weight of carrying or carrying people, which is unpredictable. For low-speed small unmanned vehicles, it is very important to study high-precision tracking control technology and overcome modeling difficulties for low-speed unmanned vehicles. At the same time, in order to overcome the random network delay caused by steering angle oscillation, a new method is proposed in [95]. An uncertain model adaptive model predictive control (UM-AMPC) algorithm is proposed, which can improve the tracking accuracy of unmanned vehicles under random network delays. In [96], the reinforcement learning method can improve tracking accuracy and stability and can complete trajectory tracking tasks. In [97], using reinforcement learning and PID control algorithm to realize trajectory tracking of mobile vehicles can reduce the computational complexity of reinforcement learning reward function and improve tracking accuracy.

The following will introduce the trajectory tracking control technology of low-speed small unmanned vehicles without application scenarios:

Low-speed small unmanned vehicles carrying human beings: Because the model of low-speed unmanned buses is unclear, it is difficult for traditional trajectory tracking methods to achieve a balance between accuracy and stability, and because the increase of passengers will lead to changes in the quality of unmanned buses, the model of unmanned buses will also change significantly. To solve this problem, Ref. [98] proposed a path tracking controller based on front axle reference fuzzy pure tracking control, and designed a feedback feedforward control algorithm for speed control, which improved the speed tracking efficiency and pure tracking stability. Because the driverless bus is a lagging and strongly coupled system, trajectory tracking control needs stronger robustness under different conditions. Ref. [13] designed an adaptive rolling window design trajectory tracking controller. In addition, the compensation model of heading error and the adaptive strategy of controller parameters are established to improve the control accuracy.

Low-speed small unmanned vehicles carrying cargo: For unmanned delivery vehicles, as they have the characteristics of changing cargo weight and are affected by nonlinear and time-varying characteristics, Ref. [99] proposes an improved predictive control scheme based on MTN, and uses BP as its learning algorithm for tracking control, which has good real-time performance, robustness, and convergence.

Special low-speed small unmanned vehicles: For special vehicles such as unmanned sweepers, accurate trajectory tracking is also very important. In [100], an error band control strategy is proposed by using the unique structure of low-speed small road sweepers, and the segmented PID control algorithm is applied to the path tracking of road sweepers. However, the PID controller has the problem of poor universality. When the working conditions change significantly, the control parameters are no longer the optimal parameters [101]. When the vehicle is in large curvature and complex working conditions, in order to overcome the influence of uncertain road surface and road surface curvature on the stability of vehicle trajectory tracking under difficult driving conditions, Ref. [102] designed a trajectory tracking controller based on model predictive control, which is used to deal with system state constraints and actuator drive constraints, and solved how to realize effective trajectory tracking control without lateral velocity. In view of the strong coupling, nonlinearity, and parameter uncertainty of vehicles, in [103], an improved sliding controller is used to track the trajectory of autonomous vehicles, so that the coupled longitudinal and lateral speeds, lateral deviations, and yaw angles converge asymptotically to a given equilibrium position, thus eliminating the influence of model disturbances and uncertainties. In order to get rid of the dependence on the mathematical model of the controlled system, the data-driven control method is studied. For the unmanned sweeper, in [94], an improved MFAC—is designed, which adopts the model-free adaptive control algorithm based on the pre-aiming deviation angle, and has high tracking accuracy and performance for the unmanned sweeper under different driving environments.

To sum up, it can be seen that for low-speed small unmanned vehicles, it is necessary not only to consider the internal factors of the vehicle in the process of vehicle trajectory tracking, such as the increase in garbage collected by garbage sweepers, the increase in the number of passengers on manned buses, and the number of express delivery vehicles loaded by express delivery vehicles, and changes in the weight of the whole vehicle; In addition, it is necessary to consider the road surface where vehicles run, because there will be complex road conditions such as uneven roads and accumulated water, which will lead to vehicle skidding and affect the stability of trajectory tracking [104], the parameters of trajectory tracking model cannot be described accurately, so it will be difficult to design a precise trajectory tracking controller. We can consider introducing a model-free adaptive control algorithm into the trajectory tracking system of small unmanned vehicles at low speed and combine the iterative learning control algorithm with other observers to study it [105,106], improving the accuracy and stability of trajectory tracking control.

7. Conclusions

By reviewing the key technologies of autonomous navigation of low-speed un-manned vehicles, this paper summarizes the key technical characteristics and existing problems in low-speed application scenarios, and expounds on the importance of low-speed unmanned vehicles under the current COVID-19 epidemic, which provides support for continuing the research on related low-speed unmanned vehicles, and puts forward the direction of technical upgrading and transformation. Low-speed driverless vehicles are playing an increasingly important role in various fields of human society, and are increasingly becoming the direction of large-scale application. As the foundation and core of unmanned driving technology, autonomous navigation technology has made continuous development and breakthrough, but there are still some technical problems. The following are some prospects for the future development of low-speed unmanned autonomous navigation technology.

(1) In complex environments, such as rain, snow, fog, dust, uneven road surface, accumulated water, and unstructured roads such as in the wild [107,108], the current autonomous navigation technology of unmanned vehicles can not solve the key technical problems such as positioning and scene perception [109], especially for low-speed small unmanned vehicles, which is more important. In the next step, the autonomous navigation technology of low-speed unmanned vehicles will develop in the direction of being suitable for various complex and changeable environments.

(2) At present, the autonomous navigation method of low-speed unmanned vehicles still needs to be equipped with more sensors, such as lidar, which is expensive. It is still necessary to further optimize the software algorithm, strengthen the fusion perception ability, and improve the autonomous navigation performance and adaptability to the environment in combination with artificial intelligence technology.

(3) At present, most low-speed unmanned driving technologies adopt single-vehicle intelligent schemes. In the future, multi-vehicle cooperative autonomous navigation can be carried out by combining technologies such as vehicle networking and vehicle-road coordination, which can reduce system costs, improve environmental awareness, and realize autonomous navigation networking and man-machine integration.