Automotive Radar Object Classification [P]
The author trained a five-classifier based on RadarScenes radar point clouds (cars, large vehicles, two-wheeled vehicles, pedestrians, and pedestrian groups), with input being the 16-dimensional histogram scanned per frame. The network uses a three-layer fully connected layer. This work is based on the paper “Histogram-based Deep Learning for Automotive Radar” and only processes single-frame scan data. To address the problem of data class imbalance, two-wheeled vehicles were merged with bicycles and motorized variants, and large vehicles with trucks, buses, and trains. Ablation experiments showed that changing the ratio of training/validation/test data had a greater impact on performance than adjusting the network structure or feature encoding. The macro F1 score increased as the number of radar detections per instance increased from 1 to 5, from 0.381 to 0.764; cars and pedestrians performed best, while two-wheeled vehicles performed worst. Future plans include trying spatial encoding schemes such as point cloud neural networks and accumulating multiple-frame scans to address sparsity issues and explore micro-Doppler effects.