Research On 3D Target Detection Method In Road Environment Based On Point Cloud Dat

3D object detection can effectively perceive the category and location information of interested objects in the environment,which is one of the most important technologies in autonomous driving.Point cloud is the most concise and intuitive representation for describing3 D objects.It contains rich spatial structure information.How to use it effectively to achieve3 D object detection is a research hotspot in recent years.Based on the KITTI data set in the road environment,this thesis focuses on the point cloud-based single-stage 3D object detection method 3DSSD.Based on cbam,inverse density weighting and average confidence,the feature extraction ability of the backbone network,the edge point contribution of the candidate point generation network and the branch interaction of the Anchor-free detection head were improved.Finally,three improved methods were merged to achieve an accurate and feasible end-to-end network framework.Specifically,the main work of this thesis are as follows:An improved backbone network of 3DSSD based on cbam is proposed in this thesis.Aiming at the problem that the backbone doesn’t distinguish the features of different channels and different spatial locations according to their importance,this thesis introduces the cbam attention,adding adaptive attention weights to channel dimensions and spatial dimensions.As a result,the cbam-Backbone can not only distinguish the meaning of different feature channels for the overall point cloud,but also distinguish their respective positions in space,which further strengthens its ability to extract detailed features.In addition,this thesis compares several attention mechanism structures horizontally.In the classification experiment with strong correlation with point cloud features,the high availability of cbam in the backbone network is proved.An improved candidate point generation network of 3DSSD based on inverse density weight is proposed in this thesis.Aiming at the problem of non-uniform density distribution of the original point cloud and feature points,this thesis designs an inverse density weight module,weighting the feature points of the input candidate point generation network.Densely distributed areas are assigned smaller weights,and sparsely distributed areas are given larger weights,balancing the feature weight of different sampling areas,improving the contribution of edge candidate center points to the point cloud features.An improved Anchor-free detection head of 3DSSD based on average confidence is proposed in this thesis.Aiming at the problem that the classification confidence score doesn’t correspond to the bounding box in the single-stage method,this thesis designs an average confidence strategy which uses the BEV projection and bilinear interpolation method,calculates the confidence that each sub-bounding box belongs to a specific part of the target,and uses the mean value as the final result,establishing the information interaction between the classification branch and the regression branch.Based on the above three improvement methods,this thesis uses 3DSSD as the baseline to construct separate embedded architectures,two-by-two combined ablation architectures,and an overall integrated architecture for experiments.Under the same experimental environment and parameter settings,the experimental results show that the improvement of this thesis significantly improves the average accuracy of Car and Cyclist.