3D object detection algorithm based on improved PointPillars

To address the common issue of insufficient detection accuracy in the field of 3D object detection from point clouds, particularly the poor performance in detecting small objects, this paper proposes an improved 3D object detection algorithm based on PointPillars. The paper introduces a convolutional module with a Spatial Reconstruction Unit (SRUCONV) designed to enhance the representation of point cloud features, and employs an improved GD-FPN mechanism during the feature fusion phase. This mechanism enhances the expression of point cloud features and the fusion of global information across different levels, thereby improving the model’s adaptability to complex scenarios and significantly boosting the efficiency of LiDAR 3D object detection. Experimental results demonstrate that the improved PointPillars algorithm exhibits superior detection performance on the KITTI dataset compared to existing baseline algorithms.