Multi scan radar object classification on RadarScenes [P]

Hello all,

I built a radar object classifier on RadarScenes, extending a prior single-scan classifier to accumulate observations over a tracked object's history instead of classifying each scan in isolation.

A single RadarScenes object instance contains only about 2.9 radar points on average, very sparse. A single scan also can't capture temporal characteristics: RCS and micro-Doppler both vary continuously as an object moves. Pedestrians produce characteristic micro-Doppler from limb motion; different object classes show different RCS fluctuation patterns as aspect angle and scattering geometry change scan to scan. Accumulating observations gives both higher point density and provides temporal dynamics.

Multi-scan baseline

DeepReflecs encoder (Ulrich, Glaser & Timm, RadarConf 2021), PointNet style, per point shared weights, on single scans across car, large_vehicle, two_wheeler, pedestrian, pedestrian_group: 0.7370 macro F1.

Using RadarScenes' persistent `track_id`, I build a causal, N=20, per track sliding-window buffer:

- x_seq/y_seq: Global, odometry-corrected coordinates recentered per scan on the object centroid. Unlike x_cc/y_cc (car-frame coordinates that accumulate over time to form a trajectory).

- Cross sensor buffer: whichever of the 4 sensors currently observe the track push to the same buffer.

- Stride 1, causal: every new scan updates the buffer and produces a prediction. No future context, real time streaming compatible.

- Each scan is encoded once by a frozen per scan encoder and cached

- Fusion concatenates the causal GRU's hidden state (order aware) with an order-invariant pooled embedding (all N scans' points as one set, no sequence structure) through a small trained mlp head.

Results

Model	Macro F1	Delta
Single scan	0.7370	(baseline)
20 scan point pooling	0.8613	+0.1243
Causal GRU	0.8895	+0.0282 over pooling
GRU + pooled embedding (fusion)	0.8897	+0.0002 over GRU, noise

Pooling alone, no sequence model, no notion of scan order at all, recovers +0.1243 macro F1. The GRU adds a real but much smaller +0.0282 on top. Fusion adds nothing measurable beyond the GRU.

Ablation

Llarger GRUs, a Transformer, a state space model, point level self attention, all trained on the exact same frozen per scan embeddings, land inside a 0.86 to 0.89 band, a 0.03 spread. End to end fine tuning of the frozen encoder makes things slightly worse (about -0.002 to -0.003), not better.

Conclusion

In this setup, the largest gain comes from giving the model more observations of the same tracked object: 20-scan point pooling improves macro F1 from 0.7370 to 0.8613 without using scan order at all.

Temporal modelling then provides a further, meaningful improvement. The causal GRU reaches 0.8895, adding +0.0282 over the pooled representation. So temporal ordering clearly contributes useful information; it just accounts for a smaller portion of the overall gain than observation accumulation.

With the per-scan encoder frozen, the different sequence architectures tested, suggests that the quality of the per-scan representation is the bottleneck than the particular mechanism used to aggregate the sequence.

Full report, every ablation, confusion matrix, coordinate frame reasoning: https://github.com/brunopinto900/radar-ml-autonomous-driving/blob/track-accumulation/final_report.md

Thank you.