Optimized 3D object detection algorithm for complex road scenes

In response to the weakness of traditional 3D point cloud object detection algorithms in detecting small targets with low accuracy, an improved PointPillars method based on spatial attention mechanism is proposed. After the formation of the pseudo-image by the pillar feature network, a GE (Gather-Excite) module is added. This module first allows the feature map to obtain spatial information through Gather, and then matches this spatial information with the input through Excite to produce attention information, using feature context to enhance the expressive power of convolutional neural networks. Finally, the improved algorithm is verified using the public dataset KITTI. The experimental results show that this method can accurately detect small targets such as pedestrians and cyclists, and also has significant effects on car categories. Compared with the benchmark model, the 3D average precision (AP) of the three target categories under medium detection difficulty conditions has been increased by 0.71%, 1.23%, and 2.15%, respectively, proving the effectiveness of the proposed improvement.