An Action Recognition Method Based on Dense Trajectories

Abstract

The invention discloses an action recognition method based on dense trajectories, which includes: 1) acquiring dense points for each frame of an input video with a dense sampling method, and tracking the dense points in the optical flow domain to form the trajectory of the video; 2) ) Screen the trajectory, extract the trajectory located in the central area as the foreground trajectory, and delete the trajectory outside the area as the background trajectory; 3) Extract the shape feature of the trajectory, the gradient direction histogram feature, the optical flow histogram feature, Moving edge histogram features, and moving neighborhood features; 4) For each feature, the enhanced local cascaded descriptor vector method is used for feature representation, and the vector representation of the five features in 3) is obtained, and the five features The vectors are concatenated to obtain the final middle-level representation of the video; 5) The support vector machine is used for feature classification to obtain the recognition accuracy.

Description

technical field [0001] The present invention relates to the field of computer vision, and more specifically, to an action recognition method based on dense trajectories. Background technique [0002] The development of science and technology has made camera equipment popular, and a huge amount of video data has also been generated. At the same time, video-oriented applications have emerged as the times require: intelligent video surveillance, video data classification, advanced human-computer interaction, etc. In these applications, understanding human actions is the core concern and the core content of people's research. [0003] Due to the great potential value of human action recognition, this topic has been a research hotspot for at least five years, and many methods have been proposed, such as: methods based on dense trajectories (DT), methods based on spatiotemporal interest points And methods based on convolutional neural networks, etc. Among them, the DT-based met...

AI Technical Summary

Problems solved by technology

The above methods have achieved good results, but the existing trajectory-based research usually ignores the relative positional relationship and motion relationship between trajectories, and these clues are helpful for improving the accuracy of trajectory-based recognition.

Images