A novel violence detection for drone surveillance system

Violence has always been a massive problem for all countries. One of the most effective ways to prevent Violence is to use surveillance systems that detect Violence. However, human-supervised Violence detection systems are ineffective; while traditional automated Violence detection technologies have many limitations, deep learning has emerged as the leading solution for automatically detecting Violence. Although several deep learning research papers give good experimental results, only a few are capable of real-time implementation, and fewer have been implemented in real-time on monitoring systems. In this paper, a solution is proposed to deploy a real-time violence detection system by deep learning on UAVs. The resolution will meet many standards simultaneously, such as calculation speed, model capacity, accuracy, high coverage, rapid deployment, and object tracking technology; especially the calculation speed, model capacity, and accuracy completely meet the requirements of real-time deployment. Experimental results can confirm that the concluded performance with 93.69% mAP(0.5), 0.114 FLOPS(B), weights size just 564.3 KB, and could achieve 21-22 FPS.