Automatic Recognition of Suspicious Human Activities from Surveillance Videos Using Deep Learning

Abstract

The increasing deployment of surveillance cameras in public and security-sensitive environments has resulted in the continuous generation of video data that is difficult to examine manually at scale. Conventional surveillance largely depends on human operators to observe camera feeds and identify events requiring attention, which can become inefficient because of prolonged monitoring, multiple video sources, and the complexity of human activities. This paper presents an automated deep-learning-based framework for recognizing human activities from surveillance videos. The proposed system uses OpenCV for video acquisition, frame extraction, and preprocessing, followed by a hybrid Convolutional Neural Network–Long Short-Term Memory (CNNLSTM) model for activity classification. The CNN component extracts spatial characteristics from individual video frames, while the LSTM component learns temporal relationships across successive frames. The implemented system considers six activity categories: Normal, Fighting, Fire, Burglary, Accident, and Shooting. Following classification, the predicted activity and associated confidence score are presented to the user, and an email notification can be generated when the detected activity requires an alert. The experimental evaluation reported for the developed framework achieves an accuracy of 96.50%, precision of 96.04%, recall of 97.00%, and F1-score of 96.52%. These results indicate that the CNN-LSTM approach can effectively learn spatial and temporal patterns from the considered surveillance-video data. The proposed framework provides an automated approach to surveillance activity analysis and can reduce dependence on continuous manual observation while supporting faster identification of potentially suspicious events.