Lightweight Hybrid Models for Anomaly Detection in Privacy-Aware Smart Surveillance: Edge-Based Real-Time Analysis

Real-time anomaly detection and object recognition play a crucial role in ensuring public safety through intelligent surveillance systems. This study enhances existing methods by integrating advanced deep learning techniques, focusing on computational efficiency and real-world deployment. A novel hybrid approach combining 3D ResNet-18 for temporal awareness and CSRNet with dilated convolutions for spatial feature extraction is proposed. Compared to previous works, our model achieves a 40% reduction in parameters while maintaining robust anomaly detection capabilities. Additionally, adaptive clip sampling optimizes frame processing rates based on crowd density predictions, improving accuracy and efficiency. Benchmark evaluations demonstrate significant performance gains, with an AUC-ROC of 0.91 and a processing speed of 24 FPS. These advancements make the proposed system suitable for deployment in diverse environments, from urban centers to resource-limited areas. Future research will focus on integrating Vision Transformers for improved contextual understanding and optimizing hardware performance for embedded systems.