Research On Fitness Movement Detection And Application Based On Improved YOLOv5s

The national strategy of national fitness is becoming more popular.This popularity has led to a growing integration between intelligent industries and the modern sports industry.As a result,the sports industry is experiencing a boost in high-quality development.In order to better align with national policies and accelerate the application of fitness action counting,this paper combines technologies such as target detection,posture estimation,and counting algorithms to construct a dataset of fitness actions.An improved YOLOv5 s model is employed to detect human information.Additionally,POSENET posture estimation is utilized to detect human skeletal key points.Based on different fitness actions,corresponding counting algorithms are designed.Finally,these technologies are incorporated into a WEB intelligent application.The specific work is as follows:(1)Building an image dataset of fitness actions by collecting data through field shooting or web crawlers,followed by image selection,annotation,enhancement,and division.(2)Improving the YOLOv5 s Model: Incorporating five enhancement modules into the original YOLOv5 s architecture.The enhancements are evaluated through a single experiment,ablation experiments,traditional experiment comparisons,and a comprehensive experimental analysis.The selected modules include the SPPFCSPC structure,CBMA attention mechanism,transposed convolution,and decoupled heads.These modules are combined to enhance the precision of the model.(3)Counting Algorithm for Fitness Movements: Comprising Human Skeleton Keypoint Detection and Counting Algorithm.The human skeleton keypoints are determined using the Pose Net pose estimation model,and different counting algorithms are employed based on specific fitness movements.(4)Implementation of an intelligent counting application for fitness actions through a WEB platform built using the front-end VUE framework and the back-end Flask framework.The experimental results demonstrate that the improved model of YOLOv5 s outperforms the original model with a 2.58% higher m AP＿0.5 index,achieving 96.89%;The m AP＿0.5 of the original POSENET model ranges from 75% to 95%,while the fused model achieves a m AP＿0.5 in the range of 87% to 93%,exhibiting more consistent detection accuracy;The counting algorithm’s error angle adjustment remains within 5°,ensuring precise counting;Moreover,the WEB intelligent counting application is user-friendly and provides strong observability.