BTQS: A Consensus-Based Temporal-Spatial Metric for Forensic Video Analysis

Object detection systems are increasingly used in long-duration surveillance applications. Continuous video streaming brings about additional technical issues with respect to conventional, image-based approaches. Frame-level metrics such as mean Average Precision (mAP) and Intersection over Union (IoU) often fail to characterize temporal consistency across extended recordings, where fragmented detections and unstable activation patterns may affect reliability. The paper presents a consensus-based temporal evaluation framework, which drives the detector behavior through temporal activity patterns instead of isolated frame predictions. The approach constrcuts a temporal consensus representation by combining detector activity signals and by extracting shared temporal segments. Building upon this representation, we introduce a Binary Temporal Quality Score (BTQS) that integrates temporal overlap, spatial agreement, boundary stability, and computational efficiency. The framework is designed to capture those temporal characteristics that might elude conventional, frame-wise evaluation methods.