CFAF-YOLOv8pose: Improved YOLOv8pose with Channel Feature Aggregation and Asymptotic Feature Fusion for Multi-person Human Pose Estimation

In computer vision research, the task of human pose estimation has been widely concerned. However, the problems of multi-person occlusion and complex motion pose in complex scenes bring great obstacles to pose estimation. Therefore, we propose a CFAF-YOLOv8pose architecture based on channel feature aggregation and asymptotic feature fusion to solve these problems. Specifically, we propose to adopt the Asymptotic Feature Fusion Module (AFF) to enable the model to retain rich semantic information during human pose feature extraction, while capturing detailed information from a lower level. Furthermore, we propose a Channel Feature Aggregation Module (CFA-OSA), which aggregates the key channel feature information extracted by the multi-branch topology network with the channel attention at one time, so that the model can realize the information flow between channel features, and at the same time improve the ability to extract channel feature information, so as to improve the accuracy of key point detection. Our method has achieved competitive results on both the self-built motion dataset and the public dataset OCHuman. Compared with the baseline YOLOv8spose, CFAF-YOLOv8pos achieves improvements of 4.5% in , 2.2% in , and 2.3% in , showing better performance.