Enhancing ADAS Safety: Pedestrian Motion Classification with mmWave Radar Point Cloud

Early and precise pedestrian behavior recognition remains a challenge for Advanced Driver-Assistance Systems (ADAS) safety. This paper evaluates two classification methods for pedestrian motion using low-resolution 4D mmWave radar in real-world urban environments. We collected a data set of radar point cloud that captures four different types of pedestrian behavior on the road: “Walk”, “cross”, “hesitate”, and “hand-up”. First, point cloud data are processed and clustered. The resulting clusters are tracked, and relevant features are extracted. The machine learning (ML) approach uses a Support Vector Machine (SVM) classifier achieving 84% accuracy. We proposed a lightweight deep learning architecture based on the MMPoint-GNN model, achieving 98% accuracy and outperforming both the baseline and ML models.