Enhanced DetNet: A New Framework for Detecting Small and Occluded 3D Objects

To mitigate the impact on detection performance caused by insufficient input information in 3D object detection based on single LiDAR data, this study designs three innovative modules based on the PointRCNN framework. Firstly, addressing the issue of the Multi-Layer Perceptron (MLP) in PointNet++ failing to effectively capture local features during the feature extraction phase, we propose the Adaptive Multilayer Perceptron (AMLP). Secondly, to prevent the problem of gradient vanishing due to the increased parameter scale and computational complexity of AMLP, we introduce the Channel Aware Residual module (CA-Res) in the feature extraction layer. Finally, in the head layer of the subsequent processing stage, we propose the Dynamic Attention Head (DA-Head) to enhance the representation of key features in the process of target detection. A series of experiments conducted on the KITTI validation set demonstrate that in complex scenarios, for the small target “Pedestrian”, our model achieves performance improvements of 2.08% and 3.46%, respectively, at the “Medium” and “Difficult” detection difficulty levels. To further validate the generalization capability of the Enhanced DetNet network, we deploy the trained model on the KITTI server and conduct a comprehensive evaluation of detection performance for the “Car”, “Pedestrian”, and “Cyclist” categories.