Research On The Algorithm Of Object Detection And Tracking From Point Clouds For Autonomous Vehicle

With the rapid development of artificial intelligence,autonomous vehicles have become a key development direction of the automotive industry.Autonomous vehicles are required to accurately perceive the road environment,and environmental perception technology is a key technology in the field of autonomous driving.One of the important parts of environmental perception technology is the use of lidar for 3D target detection and tracking.The effectiveness of 3D target detection and tracking is the core part that is directly related to vehicle safety.This paper uses lidar sensors to study 3D target detection and tracking technology.The main research contents include the following parts:Firstly,in view of the low accuracy of the 3D target detection method based on cluster segmentation and the lack of accurate pose information of the target,a target detection algorithm based on adaptive Euclidean clustering is proposed.The algorithm applies the sector box segmentation method to improve the ground segmentation module,uses the adaptive threshold method based on lidar parameters to improve the cluster segmentation module,extracts the pose information of the target obstacle box model based on the random sampling consensus algorithm,and outputs the information about location,size and orientation of the obstacles in the environment.After three kinds of lidar data sets experimental verification,it can accurately detect the pose information of road obstacles,and has high real-time performance.Compared with the minimum enclosing rectangle method and the three-point estimation method,the accuracy is improved by 8.14% and 6.57%,and the calculation time is shortened respectively by 23.81% and 24.71%.Secondly,a single-stage deep learning neural network model(MF-SSD)is constructed.The model uses the voxelization method to extract point cloud features.The model uses the voxelization method to extract point cloud features and uses 3D sparse convolution to reduce the calculation amount of the network model for shortening the calculation time.It improves the expressive ability of the network model by performing multiple adjacent fusions of feature maps with multiple levels of abstraction and uses a two-branch prediction network to output the bounding box and category information of the target respectively.Experiments are carried out on the KITTI dataset,and the results show that the detection accuracy of MF-SSD is significantly improved and the robustness is stronger.Finally,to track multiple targets in the road environment,the IMM-UKF-JPDAF multi-target tracking method is formed by using interactive multi-model,unscented Kalman filtering and joint probability data association filtering.This method uses IMM to realize the switching of the motion model under various conditions of the target for enhancing the adaptability of the motion model.And it uses UKF to deal with motion nonlinear problems by target motion estimation.It utilizes JPDAF to achieve data association for the reduction of clutter interference.Through the experiments in three experimental scenarios,the effectiveness of the multi-target tracking algorithm is verified.The results show that the multi-target tracking method performs well in all experimental scenarios.