Smart surveillance systems: a transformer-based approach for recognizing suspicious activities

Amidst the expanding landscape of the Internet of Things (IoT), integrating intelligent surveillance systems within smart environments is crucial for enhancing public safety and security. This study presents a transformer-based activity recognition (TAR) system that leverages IoT enabled surveillance to autonomously detect and localize suspicious activities in real time. The system utilizes the ResNet 50 architecture for feature extraction and the Detection Transformer (DETR) for precise activity detection. The system was developed using a custom dataset of 6684 images, collected from diverse surveillance environments and augmented to improve training. The dataset encompasses both suspicious and non-suspicious activities, ensuring broad representation. Key performance metrics, including precision, recall, mean average precision (mAP), class loss, box loss, and generalized intersection over union (GIoU), were evaluated to assess the system’s efficacy. The model achieves an accuracy of 91.07% in identifying and localizing activities such as harassment, attacks, weapon possession, fire, and theft. A comparative analysis with established models, including RetinaNet and Fast RCNN, highlighted TAR’s superior performance, confirming its effectiveness in real time surveillance. The findings demonstrate the potential of the TAR system for smart city applications, providing a scalable and efficient solution for improving security measures. This research contributes to the ongoing advancements in IoT enabled surveillance systems and emphasizes the need for standardized IoT integration in future security infrastructures.