3D Object Detection of Traffic Participants in Chinese Urban Roads Based on PV-RCNN

In intelligent transportation systems, 3D object detection provides precise data for object recognition and localization in dynamic traffic environments. However, the dynamic traffic environment in Chinese urban roads further increases the complexity of 3D object detection algorithms. This paper proposing an enhanced orientation feature extraction point-voxel region-based convolutional neural network (EOFE PV-RCNN) for 3D object detection in these environments. Using the OpenPCDet framework and the SusTechPOINTS annotation tool, a roadside perception point cloud dataset for Chinese urban road scenarios is annotated. Based on the annotated orientation values, a four-direction classifier is introduced, the additional convolutional layers for angle feature extraction are added, and batch normalization is applied. The default rectified linear unit (ReLU) is replaced with a leaky rectified linear unit (LeakyReLU) to enhance the model’s ability to capture motion orientation features. Experiment of the 3D object detection model showed that the EOFE PV-RCNN outperforms the classic PV-RCNN across all performance metrics, particularly improving orientation prediction accuracy by 11.83%.