Treffer: An Intelligent Deep Learning-Based Framework for Suspicious Criminal Activity Detection in Surveillance Systems: Design, Implementation, and Evaluation

The rapid expansion of smart cities and widespread deployment of surveillance infrastructure have highlighted the need for intelligent systems capable of detecting suspicious and criminal activities in real-time. Traditional surveillance systems primarily rely on manual monitoring, which is error-prone and unable to scale with the increasing volume of video data. To address these challenges, this paper presents a novel deep learning-based framework that integrates convolutional neural networks (CNNs) for spatial feature extraction with Transformer encoders for temporal sequence modeling. The proposed system effectively captures both short-term motion anomalies and long-term behavioral patterns, thereby enhancing detection accuracy in diverse environments. Experimental evaluations on benchmark datasets such as UCF-Crime and Avenue demonstrate significant improvements over state-of-the-art approaches across metrics including accuracy, F1 score, and AUC. Furthermore, edge deployment on NVIDIA Jetson Xavier NX confirms the framework’s viability for real-time operations, achieving sub-300 ms inference latency without compromising detection quality. The framework is modular, interpretable, and scalable, making it suitable for integration into smart city surveillance ecosystems. In addition to technical contributions, ethical considerations such as fairness, transparency, and privacy are addressed to ensure responsible deployment of automated surveillance systems.