A Deep Learning Approach for Pedestrian Behavior Interpretation Based on Radar Point Clouds

Recent developments in radar technology, such as resolution enhancement and height estimation, have opened up new opportunities to explore radar-only solutions for integral vehicle safety systems in the context of environmental sensing. In addition to localization and classification tasks, the interpretation of object behavior is also important, especially when pedestrians are involved. In this context, this work presents the possibility of using a deep learning approach for the interpretation of pedestrian behavior using 5D radar point clouds. Based on a PointNet++ architecture, different models were trained and evaluated on our own dataset, which includes different path scenarios. The final test dataset used was formed by different subjects. The best model had an accuracy of 88.69%, showing good generalization for classifying different subjects and validating the use of radar-based methods for classifying pedestrian behavior.