Research on 3D object detection for point cloud

With the development and advancement of information acquisition technology, 3D sensors can capture 3D target data with rich scale information and shape. Point clouds, as a common form of 3D data representation, retain the geometric information of the original 3D space. In harsh outdoor environments, point clouds for bad weather conditions are of great importance for target detection due to the strong interference of the associated noise points.This paper aims to analyse existing 3D target detection algorithms and discuss their advantages and limitations in order to address the problem of low accuracy of 3D target detection under outdoor severe weather conditions. The paper uses the severe weather CADC dataset and pre-processes this dataset with label evaluation, format transformation, and data encapsulation operations to improve the efficiency of point cloud data loading. Then, five different algorithms, including ImVoxelNet algorithm, PointNet++ algorithm, SECOND algorithm, 3DSSD algorithm and PointPillars algorithm, are proposed in this paper for handling severe weather. Finally, we perform an experimental evaluation on the CADC dataset. The experimental results show that these algorithms are effective in detecting 3D targets with high detection accuracy and robustness for mildly harsh environments; for non-lightly harsh environments, there are differences between the algorithms: the PointPillars algorithm has the best accuracy, followed by the 3DSSD algorithm, SECOND algorithm, PointNet++ algorithm and The ImVoxelNet algorithm was the least accurate.