Research On Near-field Obstacle Detection And Motion State Estimation Method For Intelligent Vehicle

Perceiving or understanding the environment surrounding the vehicle is an important step in driving assistance systems and for the functioning of autonomous vehicles. It is of utmost importance to be aware of other static or dynamic objects present in the environment along with the host vehicle. Broadly, environment perception includes: getting relative positions of static objects(also called map of the environment), finding position of host vehicle in this map and detecting moving objects and obtaining their properties(position, velocity, direction of movement etc.).In this paper, near-field obstacle detection and its state estimation for intelligent vehicle has been studied. And three main issues have been solved:1. Research on the near-field obstacle detection using multi single line laser radars. The cluster algorithm which is commonly used for obstacle detection is introduced. We compared the DBSCAN algorithm, K-means algorithm and Adoptive Breakpoint Detection algorithm. Then we introduced the basic theory of multi-sensor information fusion, fusion framework, fusion level and fusion algorithm. In order to improve the accuracy of obstacle detection, a new obstacle detection algorithm is proposed in this paper which using the radius vector, the change rate of radius vector and radar data number to cluster. This algorithm can distinguish between different objects through three different physical quantities and the detection result is more accurate and can distinguish the dynamic obstacle from static obstacle. In the clustering process, the normalized Euclidean distance is used to unify these three different physical variables. The obstacle detection algorithm not only can be used for single line laser radar, can also be used for multi-line laser radar. At last we put forward the obstacle detection architecture based on the feature level multi sensor data fusion algorithm.2. Research on the target motion state estimation problem. We introduced the concept of the tracking gate and the commonly used two kinds of tracking gate, rectangular tracking gate and oval tracking gate. Then we introduced four kinds of data association algorithm, the NNDA algorithm, probabilistic data association algorithm, joint probabilistic data association method and multiple hypothesis tracking method, and three filtering estimation algorithm, Kalman filter, α- β filter and particle filter, and four kinds of motion model, the uniform model, constant acceleration model, singer model and "current" statistical model. For the real-time performance, rectangle measurement gate and the NNDA algorithm have been choose. Then using the Kalman filter based on the "current” statistical model to estimate the target motion state.3. A simulation and experimental platform for the near-field obstacle detection and its motion state estimation were established. We use CARSIM-SIMULINK to test the target motion state estimation algorithm. Then the experimental verification is carried out on the Audi Q5 intelligent vehicle equipped with four HOKUYO-UTM-30 LX single line laser radars, and the near-field obstacle detection algorithm and the target motion state estimation algorithm are verified. After 20 experiments, it shows that the accuracy of the detection algorithm for dynamic obstacles can reach 100%, and the accuracy of static obstacles can reach about 90%. Adaptive Kalman filter based on "current" statistical model can accurately estimate the position, velocity and acceleration of the vehicle on the road.