Uncertainty-Guided Selective Temporal Refinement for mmWave Radar Human–Background Recognition

This study evaluates the incremental benefit of selective temporal refinement beyond Static Random Forest (RF) classification and causal Majority-3 stabilization in mmWave radar human–background recognition. Static RF uses 41-D current-frame features and its intrinsic tree-probability disagreement for routing, without an additional trained router. When disagreement exceeds a development-selected threshold and consecutive three-frame (K3) history is valid, the system invokes Temporal RF with a 54-D input combining 41-D current-frame features and a 13-D temporal extension; otherwise, Static predictions are retained. All paths use the same causal Majority-3. Across eight independent test acquisitions (16,742 frames), Static, Selective, and Always produce 86, 72, and 71 final errors. Relative to Static, Selective reduces Background false positives by 14, decreasing residual errors by 16.28% while retaining two Human misses. Selective invokes Temporal RF 479 times (2.8611% of all frames), versus 15,758 times (94.1226%) for Always, reducing additional temporal calls by 96.96%. In host inference tests, Fused Selective and Optimized Always have mean latencies of 10.642 and 11.681 ms/frame, respectively, an approximately 8.9% reduction. On two independent acquisitions at an additional logistics site, Selective reduces final errors from 66 to 56. These results show that disagreement-guided selective temporal refinement provides incremental improvement beyond static classification and causal output stabilization under the evaluated conditions, with infrequent additional temporal invocation.