3D Object Detection for LiDAR Based on Improved PV-RCNN

Addressing the current issue of poor performance in detecting small targets by lidar-based 3D object detection algorithms that combine point and voxel data, this paper proposes an improved PV-RCNN object detection algorithm named CBAM-RCNN. This algorithm introduces the Convolutional Block Attention Module (CBAM) to refine the Bird's Eye View (BEV) features after voxel compression. It assigns attention weights to the channel and spatial dimensions of the feature maps to enhance the expressive power of the features and subsequently improve the model's detection accuracy for small targets. Experiments on the KITTI dataset demonstrate that CBAM-RCNN achieves a 3.32% and 1.74% increase in mean Average Precision (mAP) for pedestrians and cyclists, respectively, compared to the baseline algorithm, verifying its effectiveness. Compared to current mainstream 3D object detection algorithms, CBAM-RCNN still exhibits better detection performance, further validating its advancement.