A real-time motion sequence localization method

Abstract

The present invention provides a real-time motion timing positioning method, including: tracking human joint points frame by frame from the deep motion sequence, calculating joint motion information, and composing the motion information into a dense joint motion matrix sorted by time and joint sequence; On the basis of this dense joint motion matrix, a bilinear quadratic interpolation algorithm is used to generate multiple time-series latitude-unified action matrices, so that the deep neural network can train sample sets of multiple time scales; the introduction of spatial pyramid pooling The layer replaces the flatten layer in the classic convolutional neural network to obtain an improved convolutional neural network that can accept any input size; it uses a long-term priority time exploration strategy for locating actions in continuous behavioral videos. The invention can solve the problem that the accuracy rate and recognition efficiency of the prior art action recognition are low, and real-time detection cannot be achieved.

Description

technical field [0001] The present invention relates to the technical field of motion sequence positioning, in particular to a real-time motion sequence positioning method. Background technique [0002] Human action recognition (HAR) has a wide range of industrial applications, such as video retrieval, video summarization, virtual reality, and human-computer interaction, etc. In recent years, with the help of deep learning technology, the accuracy of HAR has improved. However, in practice, human action videos are usually long and complex, and contain multiple distinct actions. Based on this, action timing localization (TAL) has attracted more and more attention from researchers. TAL aims to identify the start key frame (SKF) and end key frame (EKF) of each action in a long and complex video. TAL mainly Answer two questions: when the action occurs and ends, and what category the action belongs to. [0003] The traditional mainstream methods of TAL are all based on hand-cra...

AI Technical Summary

Problems solved by technology

[0006] The purpose of the present invention is to propose a "one-step" real-time action timing positioning method to solve the problem of low accuracy and recognition efficiency in action recognition in the prior art, and the problem that real-time detection cannot be achieved

Images