A Pedestrian Detection Algorithm based on Improved YOLOv8

Le résumé fourni par la source

To address the problem of false detection and missed detection caused by large-scale changes and occlusion of pedestrian targets in the video surveillance pedestrian detection scenario, an improved YOLOv8 detection algorithm is proposed. Firstly, replace some ordinary convolutions with DBB modules in the backbone network, and use multiple branches and re-parameterized features to enhance multi-scale feature information. Secondly, Feature Selection Module (FSM) is added to the feature pyramid network to alleviate the conflict between scale features and adaptively fuse scale features. Finally, by integrating spatial and channel attention to construct hybrid attention (MA) and adding it to the detection head, the model's ability to locate detection targets is improved. The experimental results show that the improved network model achieves an average accuracy of 75.146% on the CUHK Occlusion pedestrian dataset, and the model inference computation is 8.3GFLOPs, which increases the model detection accuracy with a small amount of computation.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.