Abstract
Introduction: Safety in the city has become a pressing issue in modern smart cities because of the rising rate of violence, attacks, and incidences involving weapons recorded by extensive CCTV networks. Conventional surveillance systems also have difficulties in real-time recognition, scaling, and transparency, as they are not always able to identify cases of violence on time. Methods: This paper proposes DeReV, a novel explainable deep feature-level ensemble framework that integrates two variants from each of the DenseNet, ResNet, and VGG families within a multi-backbone feature fusion architecture for multi-class violence detection in videos. The model is based on the Smart-City CCTV Violence Detection (SCVD) dataset to classify CCTV video as either normal, violent, or weaponized. Results and discussion: DeReV resulted in better performance with an accuracy of 96.15 and a macro F1-score of 0.9617, which is better than the individual baseline networks such as ResNet101 (91.67%), DenseNet169 (92.30%), and VGG16 (92.94%). The ROC analysis has shown that the AUC scores are 0.99 (normal), 0.98 (violence), and 0.98 (weaponized), whereas the Precision-Recall curves are 0.99, 0.97, and 0.95, respectively, which prove the high capability of discrimination and stability under different conditions. The interpretability and transparency of the results were achieved by using Explainable AI (XAI) methods like Grad-CAM, LIME, and Counterfactual Temporal Importance to identify salient spatial and temporal features in the predictive behaviors of the decision outcomes. The suggested DeReV framework not only offers more accurate predictions and model explainability but also supports trustworthy near real-time and citizen-centered AI surveillance, building safer and more resilient smart city ecosystems.