Research On Single Stage 3D Object Detection Based On Point Cloud

3D object detection is a popular basic research direction in the field of 3D computer vision.It is the upstream task of 3D object recognition,tracking and other tasks.It has a wide range of applications in the fields of autonomous driving,augmented reality,and robotics.Researchers have carried out a series of studies on 3D object detection methods and achieved important results,but there are still some problems to be solved.On the one hand,the problem of uneven point cloud density caused by sensor characteristics affects the network generalization ability.However,existing algorithms pay little attention to this problem.On the other hand,the detection results of 3D object single-stage detectors have too many false detections and low location accuracy.Aiming at these problems,this paper studies the point cloud-based 3D single-stage target detection algorithm.The research contents are as follows:(1)For the problem that the uneven density of point cloud affects the generalization ability of the network,a density-sensitive feature extraction backbone network named DA-3DSSD is proposed based on the classic single-stage 3D object detection method 3DSSD.Firstly,densityaware point sampling(DPS)strategy is proposed for the sampling module of the Set Abstract(SA)layer in DA-3DSSD.DPS adds point density as an arithmetic factor to the sampling algorithm to increase the sampling probability of points with small density,so that more internal points can be reserved for objects with low point density,and then more effective features can be extracted.Secondly,a center density attention(CDA)module is proposed in the Candidate Generation(CG)layer in DA-3DSSD.CDA re-weights the extracted features by using the phenomenon that the location where the target exists has higher density after the CG layer shift operation,so that the network can pay more attention to the feature channels that are more likely to have the targets.The comparative experiments and ablation experiments are performed on the KITTI dataset to verify the effectiveness of the proposed method.(2)For the problem that the detection results of the single-stage 3D object detector have too many false detections and the positioning accuracy is not high,a single-stage object detector based on multi-task assistance named MASSD(Multi-task Assistance Single Stage Detector)is proposed.Firstly,a feature extraction backbone network based on feature fusion is used in MASSD.In backbone network,the shallow detail features and high-level semantic features extracted by the network are fused as the input of the subsequent multi-task detection head.Shallow detail features provide more spatial information for the localization branch,making object localization more accurate.Secondly,a localization accuracy prediction branch is added to the detection head and a difficulty-adaptive localization accuracy prediction index HIo U(Hybrid Intersection of Union)is proposed.The scores obtained by multiplying the predicted HIo U and the classification score are used as the sorting standard of NMS and the criterion for threshold screening in the post-processing algorithm,which can effectively suppress false detection.Finally,a difficulty level prediction branch is added to the detection head to assist the network to learn information related to the complexity of the target during training.This branch does not participate in the detection process to avoid increasing time-consuming.Comparative experiments,ablation experiments and Effectiveness experiments on the KITTI dataset demonstrate that the proposed method improves the average precision and greatly reduces the number of false positives.Through the quantitative calculation and visualization of the detection results,it is found that the positions of objects are more accurate.