An Evidence–Conflict Based Uncertainty Quantification Approach for Object Detection of Autonomous Driving

Abstract

High-confidence false positives (HCFPs) in object detection show a serious threat to autonomous driving (AD) safety, as it is difficult to identify HCFPs using either confidence-thresholding or common entropy-based uncertainty measures due to the high confidence assigned. This study proposes an evidence–conflict based uncertainty quantification approach for object detection in AD—a lightweight post hoc method built on a Dirichlet evidential framework. The evidence is extracted from intermediate logits of the detector’s classification head and converted into a Dirichlet representation, from which we derive a structural evidence conflict metric (Conflict) that captures biased class-wise evidence allocation and serves as a risk score for object detection in AD scenarios. The method is verified on a Transformer-based generalized end-to-end detector and evaluated on the CODA data set. Experiment results show that the AUC and PR-AUC values of Conflict increase by about 2.5 times and 1.5 times, respectively, compared with detectors’ confidence baseline. Conflict also outperforms normalized entropy, Dirichlet entropy, and probabilistic objectness (PROB).