Anchor-free 3D point cloud target detection network with attention mechanism

Abstract

The existing anchor-free three-dimensional object detection methods exhibit low accuracy for small objects. Therefore, an object detection algorithm based on CenterNet is proposed. High-dimensional features are extracted from meshed point clouds by three-dimensional convolution. Low-dimensional spatial information and high-dimensional abstract semantic information are integrated by the two-dimensional backbone network with attention mechanism. After obtaining target coarse regression by the CenterNet head, the detection accuracy is improved by the Voxel R-CNN ROI Pooling module as a two-stage regression. The pending point cloud may be imperfect, caused by occlusions and sensor malfunction. During the network training phase, the shape sensing data enhancement scheme from SE-SSD is introduced to strengthen the detection capability of small targets.Using the KITTI dataset to evaluate the designed model, the detection accuracy of cars car (medium difficulty) is 80.60%, while that of cyclists and pedestrians is 69.71% and 59.20%, respectively. The model inference speed is 21.3 FPS. The experimental results demonstrate that the proposed method can effectively improve the capability of detecting small targets while it also has good performance with large targets similar to vehicles.