Application of Enhancing YOLOv8 Algorithm Using FasterNet Structure in Human Pose Estimation
Peng Zheng, Yuanfan Hu, Yanhao Chen · 2024
Human pose estimation techniques based on convolutional neural networks have been studied more and more extensively. Although this field is developing rapidly, there are still some important issues that have not been addressed. First, increasingly complex model structures are proposed, and this network model structure increases the research cost for developers. Second, the number of parameters and the computational overhead of the models are increasing, which will prevent the models from being deployed to some embedded devices with very limited resources. This research carried out in two aspects: (1) designing a simple and efficient network structure; (2) balancing the relationship between light weight, precision and speed. Benefit from the advantages of YOLOv8, which is highly accurate, lightweight, time-sensitive, easy to deploy, and easy to improve. This paper focuses on the performance and lightweight of YOLOv8-based HPE models. An improved model is proposed by this paper which is Faster-YOLOv8. This model uses the FasterNet to reduce the number of parameters of the YOLOv8 network effect and increase the network's speed. The Faster-YOLOv8 and YOLOv8 are compared and analyzed on homemade datasets regarding both performance and speed. The improved model has 3.18 %, 1.57 %, and 2.27 % higher mAP50, mAP50-95, and precision than the unmodified model, respectively.