Unsafe Actions Detection for Humans using YOLOv7

Unsafe action detection such as (Punching, Kicking, Pointing, Pushing-up, and Hugging) is crucial because of its applications in security and law enforcement activities and reducing violence in public places. To recognize unsafe human actions in videos, this paper uses the YOLOv7 algorithm to propose a camera-based automated system that can realize unsafe acts by humans. This technique will draw bounding boxes (red or green) around the people doing unsafe actions at the exact moment. Then, as the original other methods had only partially identified unsafe human activities in particular video frames, this work offered a better version for detecting unsafe actions and also was in real-time. The main objective of this paper is to help detect and keep an eye on any two or more persons who act out unsafe behaviors in real time. The primary problem with an intelligent surveillance system is the capacity to detect and recognize unsafe human behavior in real time in a video surveillance system. The classification performance of the behavior detection models is insufficient because the scene's complexity impacts behavior recognition for video surveillance systems. We created a deep learning model based on a YOLOv7 multiscale feature combination that reduces the effects of constant appearance changes, background clutter, and pedestrian occlusion to improve the network's processing effectiveness and address the problem of low classification accuracy of human unsafe action detection. This model effectively detects any unsafe activities. Because we obtained good results and excellent accuracy, The results of the experiment indicate that our model has a mean average precision (mAP) of 99%.