What is an automatic parking system? Analysis of automatic parking path planning and tracking technology

Parking lot status detection technology based on vision

As shown in the figure below, through the surveillance cameras installed in the parking lot, based on the deep learning algorithm, real-time detection of parking space occupancy status is achieved, and this information is uploaded to the parking lot parking space management backend server to provide parking space information for autonomous parking vehicles.

(3) Parking route planning

Path planning is an important part of APS, which is mainly achieved through control algorithms. The process is divided into three stages: starting position adjustment outside the parking space, parking into the space, and posture adjustment inside the parking space.

The starting position outside the parking space should be set with corresponding distance, position and other conditions in the control algorithm so that the vehicle position meets the parking conditions; a model should be established in the parking stage to carry out reasonable path planning; in the adjustment stage, a systematic analysis should be conducted on the position and posture of the vehicle body relative to the parking space, and a plan for adjusting the vehicle in the parking space should be formulated to ensure that the vehicle meets the conditions. The realization of the above control strategies requires a large amount of actual parking data analysis, combined with the specific parking space conditions collected by the system, combining theory with practice, in order to successfully realize path planning.

Path planning is the process where the CPU calculates the best parking path and strategy based on the current position of the car, the target position, and the surrounding environmental parameters based on the relative position of the car and the target parking space. Considering the principle of automatic parking, parking path planning generally meets the following requirements as much as possible:

1) The number of actions required to complete the parking path must be as small as possible, because the accuracy error of each action will be passed on to the next action. The more actions, the worse the accuracy.

2) During the implementation of each action, the angle of the vehicle's steering wheel (mostly the front wheel) needs to be consistent. Because the system is implemented through an embedded system , and the performance of the embedded system is limited, the consistent steering wheel angle can reduce the calculation of the motion trajectory to a geometric problem, otherwise it requires complex integration problems, which is a challenge to the performance of the embedded system.

Path planning is even more complicated. It requires a set of rigorous algorithms and modeling processes to ultimately form the best path and control strategy before entering the next execution link - path tracking. For example, the geometric path planning method can accurately plan the best parking path without collision and fully consider the constraint space based on the relative position between the vehicle and the parking space, ensuring the accuracy and safety of automatic parking. The path planning calculation is huge and the calculation process is extremely complex.

The process of reversing into a parking space with the automatic parking assistant can be divided into the following stages: measuring the length of the parking space, activating the automatic parking assistant, and parking with the help of the automatic parking assistant. Before the automatic parking assistant can provide steering assistance to the driver, it must first measure the parking space and identify the position of the vehicle relative to the parking space. Even if the automatic parking assistant is not activated, the sensors (G568 and G569) remain operational. In this way, when the vehicle is moving forward, when the speed is less than 40 km/h (parallel parking space) or less than 20 km/h (perpendicular parking space), the two sensors located at the front of the vehicle will measure all available parking spaces on both sides of the vehicle. The detection range of these two sensors is about 4.5 meters. The above method can also find and identify parking spaces on bends or curves, just like on straight roads. In addition to vehicles, the system can also detect other objects and parking spaces behind an object or between two objects. If PLA does not detect smaller objects in front of the parking space, a warning sound is emitted by the park distance control system when the vehicle approaches these objects. Regardless of whether the parking space is on the left or right side of the road, the data of the last parking space measured are temporarily stored in the control unit of the automatic parking assist system. When a new parking space is found or the vehicle has moved away from the previous parking space (more than 15 meters away from a parallel parking space and more than 8 meters away from a perpendicular parking space), the data about the previous parking space is deleted. If the automatic parking assist system is turned on by pressing the PLA button within the effective range, the parking space recorded in the control unit will be shown on the instrument cluster display as a blank area in a rectangular shadow. The following figure can help to understand the process of measuring a parking space on the right side of the road.

The parallel parking measurement is shown in the figure below. The length of a parallel parking space that meets the requirements should be greater than the length of the vehicle body plus the maneuvering distance and safety distance (at least 0.4 m left in the front and rear). The speed when passing the parking space should be less than 40 km/h. The best starting position of the vehicle should be next to the parallel parking space, in the direction of travel, and the distance between the side of the vehicle and the parked vehicles is 0.5 to 2.0 m.

Measuring a vertical parking space is shown in the figure below. The length of a vertical parking space that meets the requirements should be greater than the length of the vehicle plus the maneuvering distance and safety distance (at least 0.35m left on the left and right). The speed when passing through the parking space should be less than 20km/h. The best position of the vehicle should be next to the vertical parking space, in the direction of travel, and the distance between the side of the vehicle and the parked vehicles is 0.5~2.0m.

Generally, parallel parking and perpendicular parking use the paths shown in the figure below.

Parallel parking is divided into single and multiple times: single means parking once along the path shown in the figure below;

Multiple times means when the length of the parking space is relatively small, you can use the method of "rubbing the parking space" multiple times to park.

There are many uncontrollable factors in the parking process, such as steering system execution speed and accuracy, reference obstacle position changes, etc., which may cause the parking trajectory to deviate from the path planning trajectory. For this reason, the parking trajectory dynamic planning technology has been developed, which can realize real-time trajectory correction and even trajectory re-planning during the parking process, as shown in the following figure:

The indoor positioning technology is shown in the figure below. It solves the problem of no GPS in underground parking lots by adopting visual SLAM + tag-assisted positioning. At the same time, it improves positioning accuracy by fusing multi-source information.

Global and local path planning techniques

As shown in the figure below, global path planning between any two points is implemented based on the A* algorithm, supporting path planning reset, route selection and speed planning functions. At the same time, combined with the real-time environmental perception status, local path planning is performed to achieve autonomous decision-making such as emergency braking, cruise control, lane change avoidance, and lane change overtaking.

Automatic parking path planning based on improved deep reinforcement learning

A motion planning method based on deep reinforcement learning is proposed. The vehicle posture, steering wheel angle and minimum distance to obstacles are used as states, and the target steering wheel angle is used as the action. The parking algorithm framework based on deep reinforcement learning is built through Pytorch. A guided reward function is designed to avoid the reward sparseness problem; the average round reward is used as the priority, and the experience pool is improved to store and eliminate samples based on the priority queue; for the parking problem, a phased training method based on curriculum learning is proposed to accelerate the convergence of the algorithm. Simulation The results show that the proposed algorithm converges 25% faster than the original algorithm. The trained intelligent agent has strong planning ability and robustness, with a planning success rate of 90.6%, and has good comfort and safety.

0 Introduction

The motion planning methods used for automatic parking can be divided into rule-based methods and learning-based methods. Compared with the above planning methods, deep reinforcement learning has the advantages of strong solving ability and autonomous exploration. Many researchers and institutions use deep reinforcement learning to solve control problems and have achieved good results. Although automatic parking path planning based on deep reinforcement learning has good application prospects, there are few studies on it at the theoretical and application levels, and there are no successful cases of industrialization. If it is to be applied and promoted on a large scale, there are still problems such as low exploration efficiency and difficulty in convergence that need to be solved urgently.