Research On 3D Object Detection Algorithm Based On Point Cloud

Environmental perception is one of the key technologies in the hot research field of autonomous driving,and 3D object detection,as an important research method,plays a vital role.With the improvement of hardware computing power,lidar can acquire more and more accurate three-dimensional feature information of objects in the environment,but the scale of point cloud data is huge,and the form is different from image data.Therefore,how to process the point cloud and improve the accuracy of target recognition is the core content of this study,which has important research significance.Based on the KITTI dataset,this paper studies how to accurately detect the size,direction and position of the three targets of vehicles,pedestrians and cyclists in point cloud data.Aiming at the 3D point cloud object detection algorithm in the environment perception system,this paper proposes an improved point cloud 3D object detection algorithm based on the Point Pillars network.The main research work of this paper has the following aspects:1.Aiming at the problem of redundant ground point clouds in the KITTI dataset,the Patchwork++ ground point cloud filtering algorithm was selected as data preprocessing and embedded in the model network.Experiments have proved that the algorithm can effectively remove ground point clouds and improve the recognition accuracy of the model.Aiming at the low recognition accuracy of the model for difficult samples,the shape-aware data enhancement module is added to the model network,and it is verified by experiments that the target recognition accuracy can be further improved.2.In order to solve the problem of inaccurate angle prediction and position regression in 3D object detection,the loss function structure of the network is redesigned,and GWD(Gaussian Wasserstein Distance)and KLD(Kullback Leibler Divergence)loss function are applied to the network,effectively improving the detection accuracy and robustness of the network.3.The improved Point Pillars 3D object detection algorithm proposed in this paper has achieved excellent performance in detection accuracy.Experimental results show that the performance of the improved model in both 3D mode and bird’s-eye view mode shows better accuracy,both in detecting the category of objects and detecting difficult samples.In order to further evaluate the impact of the three improved methods on the detection performance of the Point Pillars network,multiple sets of contrastive ablation experiments were performed.The experimental results verify the superiority of the algorithm proposed in this paper in terms of detection performance.