LiDAR 3D object detection based on improved PillarNet

To address the limitations of current methods in detecting small objects, such as pedestrians and cyclists, within autonomous driving scenarios, we propose a novel 3D object detection algorithm based on an improved PillarNet architecture. These limitations often manifest as reduced recognition accuracy and increased susceptibility to false positives and negatives. The proposed method helps to capture key contextual information and spatial relationships between point cloud columns by adding a global attention mechanism to the pair. In addition, the addition of residual convolution helps to enhance the feature extraction capability of the network This comprehensive strategy expands the receptive field for point cloud features and bolsters the model’s robustness in recognizing small objects within intricate environments. The effectiveness of the proposed method is validated through evaluations on the KITTI dataset. Compared to the baseline PillarNet, our improved algorithm achieves a 2.5% increase in the average detection accuracy for the cyclist category, demonstrating the efficacy of our approach.