A Method for Estimating the Posture of Yoga Asansa Exercises Based on Lightweight OpenPose

In order to significantly reduce the computational complexity to achieve efficient pose estimation of yoga pose practice, this paper studies the pose estimation method of fitness yoga pose practice based on lightweight OpenPose. First, based on the video image of fitness yoga postures, the background model of the video image is constructed through HSV color space conversion processing. The background model is used to segment the foreground and background of yoga pose practice. A lightweight OpenPose fitness yoga pose training pose estimation model is established. This model uses a 10-layer VGG19 network to deeply study the segmented yoga pose foreground image, so as to extract the feature map of yoga pose training. These feature maps are then input into a multi-stage, two-branch convolutional neural network to further calculate the joint heat map and joint affinity domain map of fitness yoga pose exercises. In order to capture and track the joints of human motion in real time, the pose distance between frames and the maximum weight matching algorithm of a bipartite graph is used to track the pose of continuous frames of video images. Through the least squares support vector machine classifier integrated into the softmax layer of the lightweight OpenPose network model, the final pose estimation results of fitness yoga exercises are output. The experiment shows that this method can effectively realize the color space conversion from RGB to HSV of the fitness yoga pose exercise action image, and can realize the foreground and background segmentation of the fitness yoga pose exercise action image. At the same time, it can accurately identify the yoga pose exercise action posture Corresponding to different frames in a fitness yoga pose exercise action video, and the application effect is better.