Automotive Radar Object Classification [P]

Hello all, I'm a radar signal processing engineer and i trained a 5-class classifier (car, large_vehicle, two_wheeler, pedestrian, pedestrian_group) on RadarScenes radar point clouds. The input vector is a per-scan histogram (16 bins) and the network is a 3-layer MLP. The loss function is a class-weighted cross-entropy loss. This work is based on "Histogram-based Deep Learning for Automotive Radar" paper. I scoped the project to be one scan only. Accumulation of multiple scans is the next step. Data Class Imbalance: two-wheelers and large_vehicles has a low number of occurences. Aggregated Classes: two_wheeler mixes bicycles and motorized variants; large_vehicle merges trucks, buses, and trains together due to data scarcity. Sequence Bias: Long tracks of slow-moving objects can skew a particular data split velocity distribution, causing high F1 score variance across folds. Ablation studies I tried with bigger MLPs, alternative feature encodings, and different histogram binning, all moved performance less than the variation caused by changing the train/validation/test split. I measured that split sensitivity across 6 folds, keeping the same proportions. Changing the histogram to per-instance statistics (mean/median/std) slightly degraded performance. Main findings Macro F1 rises from 0.381 to 0.764 as the naturally occurring number of radar detections per instance increases from 1 to 5. I trained the model normally using all available detections, then bucketed its existing validation predictions by each instance's detection count and computed macro F1 per bucket. The classes car and pedestrian has the best performance and two_wheeler has the worst. A car is often confused as large vehicle when the car was wider than usual or had a unusually high rcs (which can happen due to multipath for example). The two_wheeler is often confused as pedestrian because their vr_compensated distributions overlap, which is the the model's single most important feature for these two cla