Human body action detection and positioning method based on space-time combination

Abstract

The invention discloses a human motion detection and positioning method based on space-time combination, and the method comprises the steps: taking an unpruned video as an input, dividing the video into a plurality of short units with the same length through data preprocessing, carrying out the sparse sampling randomly, and extracting the space-time characteristics through a double-flow convolutional neural network; Secondly, entering a space-time joint network to judge the occurrence interval of actions to obtain a group of action scoring oscillograms, inputting the action oscillograms into aGTAG network, and setting different thresholds to meet different positioning precision requirements and obtain action proposal sections with different granularities; All the action proposal sectionsdetect the types of actions through an action classifier, the time boundary of action occurrence is finely corrected through an integrity filter, and human body action detection and positioning in a complex scene are achieved. The method provided by the invention can be applied to actual scenes with serious human body shielding, changeable postures and more interference objects, and can be used for well processing activity categories with different time structures.

Description

technical field [0001] The invention belongs to the field of computer graphics and image processing, and relates to a human motion detection and positioning method based on time-space integration. Background technique [0002] Human motion detection has always been an important research topic in the field of computer vision. The ability to automatically determine the time interval of an action in a video sequence, including the start time, end time, and type of action, is crucial to our understanding and application of human behavior. . In recent years, due to the explosive growth of video data and the urgent need for intelligent video processing, time-sequence-based action detection methods have received increasing attention. Temporal behavior detection can be divided into two stages: generating time interval proposals and classifying the proposals. The current methods basically pay more attention to the decision-making of candidate proposals, while ignoring the temporal s...

AI Technical Summary

Problems solved by technology

[0003] Existing algorithms, such as SSN, TAG, and CBR, use and improve the method of generating candidate areas to classify candidate areas. To a certain extent, time-series structure information is used, and the effect is more prominent in simple and specific scenes. , but in the actual engineering scene where the human body is seriously occluded, the posture is changeable, and there are many disturbing objects, the detection results of the existing methods still need to be improved.

Images