Abstract:

To address the decline in detection accuracy of distant and small-scale objects in sparse point cloud environments due to limited feature extraction capabilities， this paper proposes a novel 3D object detection algorithm， CAResPillars， based on multiple attention mechanisms and deep residual networks. Built upon the PointPillars framework， the method introduces a Point-Channel Attention （PC-Attention） mechanism in the point cloud feature encoding stage. This dual attention design， operating at both point andchannel levels， adaptively enhances critical geometric features to mitigate the feature insufficiency problem in sparse point clouds.In the BEV feature extraction stage， a deep residual backbone network named CAResNet42， integrated with a coordinate attention mechanism， is employed to strengthen the model's ability to capture features of long-range and small-scale objects while addressing the vanishing gradient issue in deep networks. Experiments conducted on the KITTI dataset demonstrate that CAResPillars achieves notable performance gains over the baseline. In particular， the mean Average Precision （mAP） for pedestrian detection， typically involving smaller 3D objects， improves by up to 6.39%， with vehicle detection accuracy also showing consistent improvement. Ablation studies confirm the effectiveness of the proposed modules， and real-world deployment validates the algorithm's robustness and applicability in complex environments. The proposed method offers an effective and practical solution for real-world 3D object detection tasks， highlighting its potential industrial value.

Key words: 3D object detection, point cloud coding, point channel attention mechanism, residual structure, LiDAR