SUSPICIOUS ACTIVITY RECOGNITION

In the current security-focused environment, the necessity for advanced surveillance systems capable of detecting suspicious activities is more critical than ever. This project introduces an innovative approach for recognizing such activities using deep learning algorithms on video data. Utilizing the DCSASS Dataset, which includes videos across thirteen distinct classes of suspicious activities such as abuse, arson, assault, and robbery, we developed a hybrid architecture integrating ResNet50 and I3D to manage the temporal and spatial complexities inherent in video data. Our model achieved significant accuracy during training, with approximately 85% accuracy on a validation set. Enhancements included data augmentation, fine-tuning hyperparameters, and ensemble techniques, prioritizing model interpretability through class activation mapping. This project aims to provide a robust, reliable solution for autonomous surveillance systems, reducing reliance on error-prone human monitoring and enhancing real-time security measures.