Consensus-driven bounding box fusion for robust weapon detection in surveillance systems

Abstract

Object detection is central to public safety and surveillance, where reliable weapon identification is critical to mitigating security risks, and missed detections can have severe consequences. Existing ensemble methods using homogeneous architectures or standard fusion strategies do not adequately address occlusion, confidence unreliability, and contributor consensus, resulting in suboptimal recall and persistent false negatives. To address these limitations, this work presents an ensemble fusion framework combining convolutional neural network and transformer-based detectors for architectural diversity, together with a proposed Contributor Consensus Weighted Box Fusion (CC-WBF) method. The framework follows a three-stage approach: benchmarking for model selection, domain-specific pretraining, and task-focused finetuning. CC-WBF extends standard Weighted Box Fusion (WBF) through transitive clustering, weighted confidence aggregation, and contributor-aware rescaling to reduce the influence of weak or inconsistent predictions. Under matched cross-architecture detector combinations, CC-WBF increases mean recall from 0.9242 to 0.9434 on the Finetuning Dataset (FD) and from 0.7594 to 0.7871 on the Merged Military Dataset (MMD), corresponding to relative false negative rate reductions of 25.31% and 11.49%, respectively. These recall gains are statistically supported across matched detector combinations, but are accompanied by an increase in false discovery rate on both datasets. The results characterize CC-WBF as a recall-oriented fusion strategy, while the larger improvement of the complete framework reflects the combined contribution of architectural diversity and fusion. These findings support a statistically grounded, recall-oriented approach to weapon detection in surveillance, with implications for safety-critical computer vision. Implementation code and experimental resources are publicly available at https://github.com/MuhammadIshtiaq/RQ1-Dataset-and-Code.