基于深度学习的立定跳远视觉检测方法研究

Research on Visual Detection Method of Standing Long Jump Based on Deep Learning

针对现有立定跳远视觉检测方法中起跳和落地瞬间判断不够准确、落地瞬间脚部遮挡导致测距点缺失影响跳远测距精度的问题,提出一种基于深度学习的立定跳远视觉检测方法.首先提取测试者的人体骨骼关键点,利用跟踪微分器对膝关节角去噪,通过膝关节角的极大值点准确判断起跳和落地瞬间;然后通过YOLO v5目标检测和帧间差分法定位到测距点,采用卡尔曼滤波对测距点进行轨迹跟踪,预测落地瞬间缺失的测距点位置,将测距点的观测值和预测值进行卡尔曼滤波融合来提高测距点的定位精度;最后进行透视变换校正,根据测距点的融合值计算立定跳远距离.试验结果表明,该方法的立定跳远测量平均绝对误差为0.497 cm,符合立定跳远测量1 cm精度的要求.

To address the issues of inaccurate judgment during take-off and landing moments and the loss of ranging points due to foot occlusion during landing in existing visual detection methods for standing long jump,a depth-learning-based detection method is proposed.First,the key points of the human body skeleton of the tester are extracted,and the tracking differentiator is used to denoise the knee joint angle.The take-off and landing moments are accurately determined by the maximum points of the knee joint angle,then the YOLO v5 object detection and inter-frame difference method are utilized to locate the ranging points;A Kalman filter is applied to track the trajectory of the ranging point,predicting the location of the missing ranging point at the moment of landing,and perform Kalman filter fusion on the observed and predicted values of the ranging point to improve the positioning accuracy of the ranging points.Finally,perspective transformation correction is performed,and the standing long jump distance is calculated based on the fusion value of the ranging points.Experiments show that the average absolute error of standing long jump measurement using this method is 0.497 cm,which meets the 1cm accuracy requirement of standing long jump measurement.