Multi scan radar object classification on RadarScenes [P]

What happened

The author developed a radar object classifier on RadarScenes, extending a prior single-scan classifier to accumulate observations over a tracked object's history. The model uses a causal sliding-window buffer with a track_id, combining a GRU with a pooled embedding for sequence modeling and fusion. The results show that accumulating more radar returns significantly improves performance over single-scan methods, with sequence models adding minimal improvement beyond pooled representations.

Why it matters

This work is significant for AI developers and researchers in the field of autonomous driving and radar-based object classification. It highlights the importance of temporal information in radar data and suggests that improving point-level representations may be more effective than exploring complex sequence models.

• Accumulating radar observations over time improves classification performance.
• Temporal dynamics like micro-Doppler and RCS variations are critical for accurate classification.
• Sequence models like GRU and Transformer add minimal improvement beyond pooled representations.
• Improving point-level representations may be more effective than exploring complex sequence models.
• The study highlights the importance of temporal information in radar data for object classification.

The bigger picture

The practical impact of this work is the potential for improved radar-based object classification in autonomous driving systems, emphasizing the value of temporal information and suggesting that enhancing point-level representations may yield better results than complex sequence models.

The study concludes that accumulating more radar returns significantly improves classification performance, with sequence models adding minimal improvement beyond pooled representations. This suggests that enhancing point-level representations may be a more promising direction than further exploring complex sequence models.

Technical details

• DeepReflecs encoder used as the baseline model.
• Causal, N=20, per track sliding-window buffer with track_id.
• x_seq/y_seq: Global, odometry-corrected coordinates recentered per scan on the object centroid.
• Cross sensor buffer: Sensors push observations to the same buffer.
• Strided causal GRU with order-aware hidden state.
• Fusion of GRU hidden state with pooled embedding through an MLP head.
• Ablation studies on various sequence models like Transformer and state space models.

FAQ