Human behavior recognition method and device based on depth camera and basic posture

Abstract

The invention discloses a human body behavior recognition method and device based on a depth camera and basic postures, wherein the method includes: detecting images of human body behaviors through a depth camera, and extracting key point information; reconstructing a three-dimensional structure of a human body according to the input image ;Reconstruct the 3D data of the human body to identify the key poses and extract the corresponding features to obtain the time sequence of the pose feature description set; match it with the pre-trained feature description set to obtain the initial time description sequence of the basic pose; remove the initial time Describe the repeated information of the sequence to obtain the final time description sequence; input the final time description sequence into the pre-trained long-short-term memory network for human behavior recognition to obtain the recognition result. This method can reduce the interference caused by different speeds and non-standard movements of the recognition target, keep the recognition perspective unchanged, make it easier to build a database, and improve the robustness and accuracy of human behavior recognition.

Description

technical field [0001] The present invention relates to the technical field of human-computer interaction, in particular to a human behavior recognition method and device based on a depth camera and basic gestures. Background technique [0002] With the development of computer vision technology and the emergence of convolutional neural networks, the accuracy and speed of computers for static recognition of multiple types of objects have been continuously improved, and computers have a certain ability to understand the natural world. At the same time, for sequence tasks, including some recognition problems with obvious time scales such as speech recognition technology and language translation, the long short-term memory network has achieved good results, because it can store historical information in the cell unit, and over time With the increase of the scale, the network structure still preserves a relatively complete historical memory, so as to achieve a good recognition ef...

AI Technical Summary

Problems solved by technology

[0003] However, good results have not been achieved for video human behavior recognition. There are currently two mainstream methods. One is based on RGB images. In addition to using the two-dimensional information of the picture, plus the time scale, the 3D convolutional neural network is used. method to identify objects, but this type of method has the disadvantages of slow training speed and easy to be affected by the environment background, and the effect is not very ideal; another method, the main idea focuses on dimensionality reduction, and the tool used is the depth camera, because the depth camera It can capture depth information, and can better describe the spatial movement information of the human body, because the main action part is still the human body in video-based human behavior recognition. In this category, the person is first positioned, and then the key points of the person are partly Extraction, the joint movement of the human body is basically determined by the key points of the body, so the information of the key points moving with time series can ideally describe the human body's movements as sequence information

[0004] However, there are still great difficulties in understanding human behavior in videos. One of them is that the angle of view may change with the shooting position or the angle of the human body relative to the camera. For example, a kind of walking may have different angles. These are for the training set. Necessary, this creates a huge database problem, and for each movement, different people will have individual differences in the time to complete

Images