Human action recognition method and system

Abstract

The invention discloses a human body action recognition method and system. The method includes: acquiring a collected human body video image, segmenting the action sequence of the human body in the human body video image, extracting the segmented action sequence, and obtaining several time series frames; Using a preset network model to perform feature extraction on several time-series frames to obtain a feature matrix; using a preset LSTM network model to classify and identify the feature matrix to obtain a human action recognition result. The present invention uses the pre-trained depth model to fully extract the rich features of the image. Secondly, for classification tasks, the traditional machine learning algorithm is not expressive enough for complex timing problems. The long-short-term memory network (LSTM) used by the present invention is just to solve the timing problem. Useful ways. Only one camera is needed, and it has the advantages of low cost, real-time detection, and high accuracy.

Description

technical field [0001] The invention relates to the technical field of human action recognition in deep learning, in particular to a method and system for human action recognition. Background technique [0002] At present, in human action recognition technology, the following methods are usually used: [0003] (1) For contact recognition, portable wearable tool sensors are widely used to record human motion data in real time, including acceleration, angular velocity, GPS, etc., and certain mathematical statistical methods are used for recognition and judgment. [0004] (2) For non-contact recognition, the ROI of the recognition target is generally positioned first, and then the input image is extracted with artificially designed features, and then some classification algorithms (such as KNN and SVM algorithms) are used for classification. [0005] (3) In recent years, with the development of deep learning technology, based on pre-trained deep network, real-time extraction o...

AI Technical Summary

Problems solved by technology

[0007] (1) Taking tennis players as an example, when tennis players are exercising, wearable devices will cause a kind of inconvenience to the athletes, and the attached sensors are relatively expensive compared to ordinary cameras

[0008] (2) An action of a tennis player is a continuous time series of multiple frames, and a single frame picture cannot judge the player's action, which makes the feature extraction of artificial design have certain limitations.

[0009] (3) The 2D coordinate information of human joint points lacks richness as an action feature, and cannot fully represent the entire action process of the player, and there is a loss of feature information

Images