SGPrior: Test Prioritization for LiDAR-Based Multi-Object Tracking Systems

Faults in the three-dimensional detection and multi-object tracking components of autonomous driving systems (ADSs) often manifest as cross-frame anomalies that single-frame confidence or neuron-coverage signals cannot adequately characterize. This paper proposes SGPrior, an offline method for prioritizing light detection and ranging (LiDAR) test sequences through scene-graph reasoning. SGPrior converts frame-wise detection and tracking outputs into attributed scene graphs. A finite-state machine activates consistency assertions whose operators construct an assertion graph for comparison with the current ADS-output graph. Normalized Hamming distances between aligned graph encodings are aggregated into a sequence-level priority score. Experiments on KITTI, nuScenes, and Waymo show that, under a 30% test-analysis budget, SGPrior selects subsets with lower multiple object tracking accuracy (MOTA), multiple object tracking precision (MOTP), and higher order tracking accuracy (HOTA), and more false positives and false negatives than the baselines. It achieves an average percentage of faults detected (APFD) of 0.957–0.958 with substantially lower cost than neuron-coverage ranking. Retraining the detectors in six detector–tracker combinations with SGPrior-selected sequences also outperforms equal-size random selection on held-out data. These results suggest that, after perception inference, SGPrior can help testers allocate a limited inspection or annotation budget to sequences that expose more detection and tracking errors. The benefit concerns the selection of sequences for analysis, rather than a reduction in the inference required to process the candidate pool.