Abstract:  The 3D object detection method based on radar point cloud effectively solves the problem that the 2D object detection based on RGB images is easily affected by such factors as light and weather. However, due to such issues as radar resolution and scanning distance, the point clouds collected by lidar are often sparse, which will undermine the accuracy of 3D object detection. To address this problem, an object detection algorithm fused with sparse point cloud completion was proposed. A point cloud completion network was constructed using encoding and decoding mechanisms. A complete dense point cloud was generated from the input partial sparse point cloud. According to the characteristics of the cascade decoder method, a new composite loss function was defined. In addition to the loss in the original folding-based decoder stage, the compound loss function also added the loss in the fully connected decoder stage to ensure that the total error of the decoder network was minimized. Thus, the point cloud completion network could generate dense points with more complete information Ydetail, and apply the completed point cloud to the 3D object detection task. Experimental results show that the proposed algorithm can well complete the sparse car point cloud in the KITTI data set, and effectively improve the accuracy of object detection, especially for the data of moderate and high difficulty, with the improvement of 6.81% and 9.29%, respectively. 

Key words:  , object detection, radar point clouds, point cloud completion, compound loss function, KITTI