Real-time human motion posture correction method and system

Abstract

The invention discloses a real-time human motion posture correction method and system. The method comprises the following steps: acquiring depth map data by using Kinect equipment; obtaining bone data from the depth map data, and carrying out stabilization and smoothing processing; detecting and identifying the human body posture through a preset method; evaluating the standard degree of the posture by calculating the posture matching accuracy; and feeding back a result in the posture changing correction process to a user in real time. The beneficial effects are that the user can select the fitness time and place at will, which is more suitable for modern busy office workers and young people; action done by a user can be standardized, damage to the body of the user is avoided, and a better fitness effect can be achieved when the mobile phone APP is used for action guidance; it can be avoided that wearing is cumbersome, and better user experience and effect can be obtained.

Description

technical field [0001] The present invention relates to the field of public fitness services based on depth-of-field images, in particular to a real-time human motion posture correction method and system, especially to a real-time human motion posture correction method and system based on Kinect joint extraction. Background technique [0002] People expect to get professional technical guidance and proper fitness effects when they are exercising. The essential requirement is to be able to make standardized movements. At present, there are three main methods: hiring personal trainers, App guidance, and smart wearable devices. [0003] (1) Hiring a private trainer is the most effective method, but due to the characteristics of the private trainer profession itself, there are 50% vacancies in the market in 2017, and because of this method, it is impossible to flexibly carry out fitness by yourself, and the cost is high, which is not suitable for modern Busy office workers, yout...

AI Technical Summary

Problems solved by technology

[0003] (1) Hiring a private trainer is the most effective method, but due to the characteristics of the private trainer profession itself, there are 50% vacancies in the market in 2017, and because of this method, it is impossible to flexibly carry out fitness by yourself, and the cost is high, which is not suitable for modern Busy office workers, youth groups;

[0004] (2) The mobile client app is more widely used than the above ones, but due to the lack of supervision and correction functions, users may not necessarily make standard movements to achieve corresponding fitness effects, and what’s more, it may damage the body;

[0005] (3) Wearable smart devices are emerging products, but they are inevitably cumbersome to wear, whether it is the low data dimension of a single device (such as a smart watch) and poor reliability, or the bloatedness of a conjoined device (such as a smart fitness suit), Both reduce the user experience and fitness effect to a certain extent

[0008] Another mainstream somatosensory device on the market, Intel's RealSense, cannot meet the high research and application requirements for tracking, identification and real-time performance during fitness, and cannot establish a full-platform image cloud library for machine learning

[0009] Among the algorithms in this aspect: 1. Conventional global method recognition (pre-segmentation model), the amount of calculation is large and it is difficult to process in real time

2. Local feature recognition based on the point fingerprint method, "Research on 3D Object Recognition Algorithm Based on RealSense" is a method of using affine transformation coefficients to process corresponding points between objects and scenes to determine the 3D coordinates of nodes, which is difficult to eliminate data jitter and relatively robust Difference

[0011] Existing human block pose recognition based on depth-of-field images, including image processing, contour matching and eigenvalue matching recognition, cannot meet the requirements of accuracy and real-time at the same time

1. In "Human Action Recognition in Video Using Kinematic Features and Multi-instance Learning", the depth image of Kinect is segmented with double thresholds, and the hand image is obtained combined with morphology to recognize gestures in real time. The algorithm is complex and the amount of parallel processing is large

2. Algorithms based on contour matching, "Research and Implementation of Kinect-Based Air Gesture Tracking and Recognition" uses color and depth images to locate joint positions and recognize human movements using bone maps obtained from color and depth images. The same amount of data calculation is too large to be real-time

[0013] In the single-frame image pose estimation method with the largest coverage of consistency, and the normalized matching algorithm based on Kinect skeleton data proposed in "Human Action Pose Method Based on Skeleton Data", the matching accuracy rate is about 90%, and it is difficult to accurately compare micro-poses Changes, without considering dynamic real-time and accuracy

Images