A Deep Learning-Driven Framework for Automated Real-Time Detection and Multiclass Classification of Suspicious Human Activities in Surveillance Video Streams

Abstract

The increasing demand for intelligent surveillance systems has motivated the development of automated methods capable of detecting and classifying abnormal human behaviors in real time. This paper presents a deep learning-driven framework for the recognition of suspicious activities in surveillance video streams. Leveraging convolutional and temporal modeling techniques, the proposed system extracts robust spatial-temporal features to accurately classify both normal and abnormal behaviors such as loitering, fighting, theft, and vandalism. Publicly available benchmark datasets, including UCF-Crime and Avenue, were used to evaluate the system, providing a diverse range of real-world scenarios for comprehensive performance assessment. Experimental results demonstrate that the proposed framework outperforms conventional baselines such as CNN+LSTM, 3D-CNN, and handcrafted feature-based methods in terms of accuracy, precision, recall, and F1-score. Furthermore, the model achieves near real-time performance with an inference speed of approximately 30 fps, highlighting its suitability for practical deployment in large-scale monitoring systems. These findings confirm the effectiveness and scalability of the framework as a reliable solution for enhancing public safety through intelligent surveillance.