CR-Pillars: A Three-Dimensional Object Detection Model Based on Enhanced PointPillars

Aiming at the issue of weak feature extraction capability of 3D object detection algorithms for onboard LiDAR in automated driving scenarios. We propose a novel CR-Pillars that employs the Convolutional Block Attention Module (CBAM) fusion attention mechanism and a custom residual block ResNet64 network to optimize the backbone network of PointPillars 3D object detection algorithm. Firstly, the point cloud representation is performed using Pillars and a pseudo image is generated. Then, the CBAM attention mechanism, which integrates channel attention modules and spatial attention modules, is utilized to improve the feature extraction of the pseudo-image. Moreover, the output of the attention module is downsampled and concatenated by the residual network, enhancing the ability of the algorithm to extract contextual semantics and global features. The Single Shot Multibox Detector (SSD) algorithm is used for 3D object regression. Experimental results on KITTI demonstrate that the proposed method achieves higher detection accuracy compared to the original method, with an average improvement of 2.1% in detection accuracy, thus verifying the effectiveness of our approach.