Abstract:Car navigation is a key technology of unmanned driving,and its main technical difficulty is dynamic obstacle avoidance,which means real time target recognition and feedback control.In order to obtain the real time movement status of the targets around the vehicle,a three dimensional point based algorithm is proposed to replace the rectangular coordinate system with a polar coordinate system,and directly establish a mapping relationship with the vehicle mounted lidar.Then,the accurate positioning of the characterizable target features is completed through the limited calculation of the area frame in the area recommendation network technology.The experiment is validated using point cloud data in the open source database KITTI to compare the point cloud features between dynamic targets (cars,motorcycles,pedestrians) and static targets (trees,buildings).The recognition results are compared with two commonly used 3D target recognition algorithms.The results show that the optimal average accuracy of this algorithm is 88.45%,and the optimal average direction similarity is 93.27%.The proposed algorithm has a better effect on the recognition of dynamic targets compared with the commonly used SECOND algorithm.Its feasibility is verified,and it has certain advantages in the field of vehicle navigation target recognition.