Human body motion recognition method based on three-dimensional bone information

Abstract

The invention discloses a human body motion recognition method based on three-dimensional bone information. The method comprises the steps that color data stream, depth data stream and bone data stream of a number of individual samples of different sex and height are processed, wherein the individual samples make various motions; SVM models corresponding to all motions are constructed; the bone data stream of a recognized target which makes any motion within the acquisition range of a Kinect depth sensor is acquired; through the bone data stream, the distance between a normalized human body bone articulation point and a reference point and 14 vector angles are acquired; the data are respectively input to multiple SVM models; and a motion corresponding to the SVM model with the highest output probability is the motion of a recognition target. According to the invention, the Kinect depth sensor is used to acquire images; the method is less affected by light condition, shadow and other factors; the depth map and the bone information of a human body motion are acquired in real time; and a human body target in a scene can be accurately located.

Description

technical field [0001] The invention relates to the field of human action recognition, in particular to a human action recognition method based on three-dimensional skeleton information. Background technique [0002] The main task of action recognition is to extract the feature data representing different actions from the depth map, and when different human bodies express the same action, there are differences in shape, clothing, and exercise habits. Therefore, how to select an effective action feature description is one of the key issues in action recognition. Since human body movements can be simply divided into upper limb movements, lower limb movements, and torso movements, and the self-occlusion problem exists during the human body's own movement process, this makes the extracted feature data redundant, which will also affect the subsequent action recognition results. Therefore, it is particularly important to select appropriate action feature data to distinguish diffe...

AI Technical Summary

Problems solved by technology

[0006] Most of the traditional action recognition methods are human action recognition based on two-dimensional image sequences, which are easily affected by lighting conditions, occlusion and shadows, and can only be obtained when the color of the human body clothes is very different from the background color or there is no occlusion. better recognition result

Although some depth cameras can obtain three-dimensional information, the effective distance of the stereo camera only depends on the baseline setting and the quality of ambient light in the scene. The TOF camera relies on the reflection of light, and its effective distance depends on the range of light emission and acceptance, and High equipment cost

Images