Radar Point Cloud Target Classification Based on Graph Neural Network

The decision-making process of the auto drive system deeply depends on the perception of the surrounding environment. At present, lidar is widely used in environmental perception, but its performance will be significantly affected in severe weather conditions (such as rain, fog, snow, etc.), which limits the reliability and safety of the auto drive system. The current mainstream solution is to use millimeter wave radar to assist decision-making and enhance perception capabilities in complex environments. As the millimeter-wave radar point cloud is sparse compared to that of LIDAR, which makes the information difficult to extract. In this paper, a new model for extracting sparse radar point cloud information is improved based on GNN networks. The model realizes the extraction of target information in the radar point cloud by using the information of the points themselves and the relationship between the points, and then combines with the detector head to realize the classification of targets. When classifying targets, using GAT detection heads instead of traditional MLP perception heads can improve the weighted extraction of different targets. A FocalLoss loss function is proposed to address the issue of class imbalance in training data, which solves the problem of unstable model results caused by class imbalance during training. Referencing the use of NLNN structure to accelerate the aggregation of node information and achieve pooling like operations to solve the problem of difficult association of information between distant points. The model achieves competitive results when evaluated on the RadarScenes dataset.