Abstract : Navigation and localization are prerequisites for intelligent vehicles in residential communities to perform their intended functions according to expectations, and this is also a core and crucial part of the programming. This paper mainly studies several methods for navigation and localization of intelligent vehicles in residential communities.
Keywords : Smart car in residential community; navigation; positioning
0. Introduction
Navigation and positioning systems are one of the key technologies in the field of intelligent vehicles, and their quality has a significant impact on the performance of intelligent vehicles. Although many research institutions both domestically and internationally have conducted in-depth research on autonomous vehicles, no car has yet achieved the same performance as a human-driven vehicle. One reason for this is that the navigation and positioning systems of autonomous vehicles have not yet met the required standards. Therefore, in-depth research on navigation and positioning systems for autonomous vehicles is not only of theoretical significance but also of practical significance.
1. Navigation Method Classification
Smart cars in residential communities have multiple navigation methods, which can be divided into three categories based on factors such as the completeness of environmental information and the type of navigation instruction signals: map-based navigation, map-based navigation, and mapless navigation.
Map-based navigation relies entirely on complete information about the environment, such as geometric models or topological maps created by the user and stored inside the smart car in the community. Based on a pre-planned global route, it uses path tracking and obstacle avoidance technologies to achieve robot navigation, as shown in Figure 1.
Figure 1. Smart car navigation map of the community
Map-based navigation uses sensors (such as odometers, sonar, laser rangefinders, visual sensors, etc.) to create a geometric or topological model map of the current environment (as shown in Figure 2), and then uses these models to achieve navigation.
Figure 2 Smart Car Map Model of the Community
Mapless navigation is a method that uses cameras or other sensors (such as ultrasonic or laser rangefinders) to detect the surrounding environment when environmental information is completely unknown. It then uses the detection of objects to identify or track them, thus enabling smart car navigation in the community.
When a smart car in a residential area doesn't fully understand its surroundings, a landmark-based navigation strategy can be employed. This involves storing distinctive features of the environment within the car. The car determines its position by detecting these landmarks and breaks down the entire route into segments between them. Navigation is then accomplished through a series of landmark detections and guidance. In relatively regular environments, a clear path marker can be drawn on the road or roadside. As the car travels, sensors continuously detect and adjust the path, ensuring the markers align with the route. When encountering obstacles, the car either stops and waits or avoids them, then returns to its original route following the markers, ultimately reaching its destination.
Since the working area and path of the intelligent vehicle in the community are fixed, map-based navigation is the most suitable among the various navigation methods mentioned above. After the working area of the intelligent vehicle in the community is determined, a complete geometric map model is created and stored in the memory of the control system, and navigation is realized based on the map. When the intelligent vehicle encounters an obstacle, it uses the automatic obstacle avoidance function to bypass the obstacle or waits for the obstacle to leave before returning to the original path.
2. Classification of Positioning Methods
Based on the sensors used, the positioning technology for intelligent vehicles in residential communities can be divided into two categories: absolute positioning technology and relative positioning technology. Relative positioning technologies mainly include ranging methods and inertial navigation methods. Among absolute positioning technologies, the more mature ones include GPS (Global Positioning System), scene recognition positioning, and base station positioning. Below is an analysis of several positioning technologies that can be referenced for intelligent vehicles in residential communities:
CellID-based positioning: Smart cars in a residential area determine their location by receiving building identification codes (CellIDs) from nearby buildings and translating this information into latitude and longitude coordinates. Disadvantage: Requires the installation of wireless network transmitters on buildings, resulting in higher costs.
GPS positioning: Tracking and guiding controlled objects in non-fixed-road systems using the Global Positioning System. Advantages: Suitable for long-distance outdoor tracking and guidance. Disadvantages: Accuracy depends on the accuracy of the GPS and the surrounding environment of the controlled object; the positioning accuracy of civilian GPS is about 10 meters, which cannot meet the requirements for use in intelligent vehicles in residential areas.
Scene recognition and localization: Images of the surrounding environment of a smart car in a residential area are generated using a CCD camera and matched with features from an environmental map stored in a computer system to determine the car's current location. Advantages: It does not require manually setting any physical paths, thus offering high flexibility; with the rapid development of computer image acquisition, storage, and processing technologies, this method is becoming increasingly practical, and its poor real-time performance can be compensated for by using high-speed processors.
Each of the different positioning methods used for smart cars in residential communities has its advantages and disadvantages. Different positioning methods can be selected based on different functional requirements and working environments. Using GPS alone for positioning has certain errors in determining the baseline. Machine vision, when performing this task, can extract feature information of the current environment in real time, improving positioning accuracy. However, using scene recognition positioning alone involves complex image processing algorithms and high positioning difficulty. Therefore, a combination of GPS and scene recognition positioning methods is considered. First, the GPS positioning system is used to narrow down the location of the smart car in the community to a range of 10 meters. Then, machine vision technology is used to find feature points in the current environment and match them with the environment map stored in the system to determine the accurate location of the smart car in the community.
3. Machine vision technology and machine vision positioning
Humans perceive the external world primarily through sensory organs such as vision, touch, hearing, and smell, with approximately 80% of information acquired through vision. Therefore, vision plays a crucial role in the entire information acquisition system. A machine vision system is an intelligent system that simulates the function of human vision to acquire and process information. A structural diagram of a machine vision system is shown in Figure 3.
Figure 3 Machine vision system
Machine vision can be viewed as the process of extracting, describing, and interpreting information from images of a three-dimensional environment. It can be divided into several main parts: image acquisition, binarization, storage, transformation, encoding, segmentation, feature extraction, image database establishment, image classification and representation, image recognition, model matching, intrinsic interpretation and understanding, etc. Based on the complexity of the methods and technologies involved in implementing these processes, they can be categorized into three processing levels: low-level vision processing, mid-level vision processing, and high-level vision processing. Although there are no clear boundaries between these levels, this classification provides a useful structure for categorizing the inherent processing procedures of robot vision systems. The implementation of vision-based robot navigation and localization requires equipping robots with vision systems.
The general working principle of a robot vision system is as follows: A CCD camera installed on a smart car in the community collects image information of the road ahead in real time, converts the reflected light intensity signal of the current road into a corresponding analog electrical signal and sends it to a video capture card. The analog-to-digital converter in the video capture card samples and quantizes the analog signal, and further converts the analog signal into a digital signal that the controller can accept and understand. The controller uses the corresponding algorithm to process the road image, identify road information and obstacle information in the road, and autonomously decides the current forward movement mode (lane change, steering) and controls the vehicle's own motion state.
In scene-based localization, images of the surrounding environment of a smart car in a residential area are generated using a CCD camera and matched with the environmental map stored in the computer system. Since overall template matching or feature extraction is computationally too intensive, boundary extraction is typically performed first during image processing, followed by further description and understanding based on image segmentation. Applying a Hough transform to the target based on the global characteristics of the image can extract boundary features. The basic principle is based on duality: calculating lines in parameter space from points in image space, and then calculating lines in image space from the intersections of these lines. This can be used to detect various curves or contours that can be represented analytically by f(x, c) = 0 (where x is the coordinate vector of an image point, and c is the parameter vector).
A representative work on machine vision navigation under structured conditions is the probabilistic reasoning-based localization method studied by Akio Kosaka and Avi Kak. This method pre-builds a wireframe model of the environment scene, uses a Gaussian distribution to describe the robot's pose vectors (position and orientation), and updates the pose during robot movement as updates to the mathematical expectation and variance of the pose vector distribution. The basic principle is shown in Figure 4, and the main steps of the algorithm are as follows:
Figure 4. Self-localization algorithm based on environment model and probabilistic reasoning
Step 1 : Describe the robot position p=(x, y, φ) using a Gaussian distribution function. After executing the motion command, the position is p′=h(P). Determine the mean P and the variance matrix Σp of the uncertainty parameter transformation for each positioning point.
Step 2 : Determine the robot's position parameters and position uncertainty distribution matrix after executing the motion command. Position uncertainty refers to the additional translational or rotational distance of the robot caused by sliding or other uncertain factors during motion.
Step 3 : Update the current position using Kalman filtering based on the scene model.
Due to the complexity of the working environment, the uncertainty of its own state, and the limitation of a single sensor only being able to obtain partial information about environmental characteristics, it is difficult to perceive the external environment by relying on only one sensor. To achieve autonomy in complex, dynamic, and uncertain environments, robots are usually equipped with multiple sensors for navigation. How to effectively utilize the information provided by multiple sensors and apply it to navigation decisions requires information fusion technology.
4. Design of a Smart Car Navigation and Positioning System for the Community
The community's intelligent vehicle navigation and positioning system mainly consists of two parts: a GPS component and a machine vision component, as shown in Figure 5. The GPS component primarily provides the absolute position coordinates, heading angle, and speed of the intelligent vehicles within the community; the machine vision component preprocesses the acquired images to obtain the navigation route.
Figure 5 Overall design of the integrated navigation and positioning system
The relative position coordinates of known points along the path are determined. After processing in both parts, the two sets of information are unified into the same coordinate system and UKF (unscented Kalman filter) filtering is performed to obtain new position information.
5. Geometric Model of Navigation Algorithm
After determining the location of the smart car in the community, the integrated navigation system matches it with the map stored in the system to calculate whether the current location of the smart car is on the path in the stored map. If it is not on the planned path, it finds the point on the map closest to the smart car and uses it as the target point, as shown in Figure 6(a). If it is on the pre-planned path, it proceeds to the next target point according to the map, as shown in Figure 6(b). Two coordinate systems are created in the navigation system: a visual coordinate system with the camera as the origin and a world coordinate system (map coordinates stored internally by the control system). The origin of the visual coordinate system is the projection point of the camera's optical center on the ground, where the positive direction of the X-axis is the negative direction of the driving direction, and the positive direction of the Y-axis is perpendicular to the X-axis and horizontally to the right. The target point in Figure 7 is the position that the cleaning robot needs to reach next. After the location of the smart car in the community is determined, the points in the world coordinate system can be converted to the location with the smart car as the primary coordinate.
Figure 6. Schematic diagram for determining the target point
Figure 7. Coordinate system diagram of the navigation system
In the visual coordinate system, its coordinates are defined as follows. The position of the camera in the geodetic coordinate system is obtained in real time by the positioning system, and its coordinates in the world coordinate system are ( ), and the coordinates of the target point in the world coordinate system are . The coordinates of the target point in the visual coordinate system are obtained by formula (1):
In the formula, represents the heading angle of the navigation vehicle. The direction of the vector is the direction of movement of the intelligent vehicle in the community.
6. Conclusion
Currently, there are various design methods for navigation and positioning systems of autonomous vehicles, but none of them are universally applicable. Different control objects and operating conditions place different demands on navigation and positioning systems, and the special environment of residential roads presents new requirements for navigation and positioning systems of autonomous vehicles. The research results of this study are of positive significance for the further research and practical application of autonomous vehicles.
References:
[1] Zhang Pengfei, He Kezhong, Ouyang Zhengzhu, et al . Multifunctional outdoor intelligent mobile robot experimental platform - THMR-V[J] . Robot, 2002, 24(2): 97-101 .
[2] Liang Xifeng . Path planning for indoor floor cleaning robots [J] . Journal of China Jiliang University , 2006 , 17(1) : 64-68 .
[3] Zou Yuanjie, Liu Xiaomei . Research on intelligent vehicle positioning and navigation system [J] . Hunan Agricultural Machinery , 2010 , 37(5) .
[4] Zhang Chao . Research on positioning of intelligent vehicle positioning and navigation system [D] . Chengdu: Master's thesis of University of Electronic Science and Technology of China, 2005 .
[5] Wei Caicai . Research on vehicle-mounted device for intelligent vehicle positioning and navigation system [J]. Power Technology Application , 2013 , (8) .
