Explainable Artificial Intelligence and Computer Vision for RealTime Detection and Prediction of Critical Events in Smart Cities

Abstract

In order to promote public safety and quick emergency response, smart cities are depending more and more on networked cameras, Internet-of-things sensors, unmanned aerial vehicles, and edge computing. However, occlusion, illumination fluctuation, camera motion, crowd density, complex relationships, and the temporal evolution of anomalous behaviour make it challenging to identify and forecast important events from continuous urban footage. For real-time critical-event monitoring, this study presents XAI-CityVision, an explainable artificial intelligence platform that combines computer vision, object identification, temporal learning, risk assessment, and visual explanation. The framework uses a synchronised acquisition layer to receive heterogeneous urban observations, preprocesses videos, uses CNN or Vision Transformer backbones to extract spatial representations, uses a YOLO-family detector to identify pertinent objects, and uses ConvLSTM or transformer-based temporal learning to model event evolution. Grad-CAM, SHAP, and attention visualisation offer complementary explanations of the choice, while a risk module integrates event likelihood, temporal persistence, and contextual severity. A universal event ontology including normal activity, accident/collision, fire/smoke, crowd anomaly, intrusion/suspicious behaviour, and traffic-related occurrences is mapped to dataset-specific labels in the experimental design, which employs public urban and surveillance benchmarks. Precision, recall, F1-score, IoU, mean average precision, ROCAUC, false-alarm rate, inference delay, and frames per second are used to assess performance. The contributions of spatial learning, object detection, temporal modelling, and multimodal context are measured by ablation studies. In smart-city settings, the suggested framework offers a repeatable architecture for integrating deployment-aware evaluation, operator-oriented explanations, and predictive performance.