EASAD: efficient and accurate suspicious activity detection using deep learning model for IoT-based video surveillance

Video surveillance on Internet of Things (IoT) devices is crucial for ensuring security and monitoring in various environments. However, these devices often operate under significant resource constraints, such as limited processing power, memory, and energy supply. As a result, traditional video analysis approaches struggle to maintain both efficiency and accuracy on IoT devices. This study addresses these limitations by proposing an innovative solution specifically designed for intelligent suspicious activity detection on resource-constrained IoT devices. Our model integrates an enhanced SqueezeNet architecture and an improved U-Net segmentation process to achieve optimal resource efficiency without compromising accuracy. The enhanced SqueezeNet architecture enables smooth operation within the computational constraints of IoT environments, while the refined U-Net segmentation process facilitates effective feature extraction for accurate analysis. Furthermore, selective feature extraction techniques, including SLBT, BoVW, and MoBSIFT, are employed to alleviate computational burden, ensuring efficient processing. These features serve as input to the resource-efficient SqueezeNet model. The model classified the suspicious actions with 95% accuracy using 90% of training data. This work represents a significant advancement in overcoming the challenges associated with video surveillance on IoT devices, highlighting the critical need for solutions that balance efficiency and accuracy in this context. © Bharati Vidyapeeth's Institute of Computer Applications and Management 2024.