A novel YOLOv8 architecture for human activity recognition of occluded pedestrians

Abstract

Perception is difficult in video surveillance applications because of the presence of dynamic objects and constant environmental changes. This problem worsens when bad weather, including snow, rain, fog, dark nights, and bright daylight, interferes with the quality of perception. The proposed work aims to enhance the accuracy of camera-based perception for human activity detection in video surveillance during adverse weather conditions. To identify primary human activities, including walking on the road during severe weather, transfer learning from many adverse conditions using real-time images or videos has been proposed as an improvement for you look only once v8 (YOLOv8)-based human activity recognition in poor weather conditions. We collected and sorted training rates into frames from videos depicting human walking activity, their combined forms, and other subgroups, such as running and standing, based on their characteristics. The assessment of the detection efficiency of the previously described images and subgroups led to a comparison of the training weights. The use of real-time activity images for training greatly enhanced the detection performance when comparing the proposed test results to the existing YOLO base weights. Furthermore, a notable improvement in human activity efficiency was obtained by utilizing extra images and feature-related combinations of data techniques.