Dense trajectory covariance descriptor-based behavior recognition method

Abstract

The invention discloses a dense trajectory covariance descriptor-based behavior recognition method, and mainly aims at solving the problem that the behavior recognition correctness is low as the prior art does not consider the correlation between different features and cannot correctly describe the movements of behavioral agents. The method comprises the following steps of: 1) extracting dense trajectories of a video, and for each pixel point in a trajectory cube, obtaining a gradient, a spatial position and time derivatives of the gradient, a light stream and a movement boundary, and taking the features as bottom-layer features; 2) obtaining a bottom-layer feature set, solving a covariance matrix of the bottom-layer feature set and projecting the bottom-layer feature set to an Euclidean space to obtain descriptors of trajectory sub-blocks; 3) connecting the descriptors of the trajectory sub-blocks in series so as to obtain a dense trajectory-based covariance matrix descriptor; and 4) carrying out BOW coding on the covariance matrix descriptor and then carrying out behavior recognition on the covariance matrix descriptor by utilizing a linear SVM classification model. The method has the effect improving the behavior description ability and the recognition correctness, and can be used for the complicated environment of video monitoring.

Description

technical field [0001] The invention belongs to the technical field of video processing, and in particular relates to a behavior recognition method, which can be used to describe video behavior in complex video monitoring environments. Background technique [0002] In the field of action recognition, artificial local features have become an effective way of feature representation. Local features do not require specific algorithms to detect human body parts, and are robust to effects such as complex backgrounds, illumination changes, and video noise. [0003] Typical local features include: spatio-temporal interest points STIP, cubes, and dense trajectories, often combined with gradient orientation histogram HOG, optical flow orientation histogram HOF, 3D gradient orientation histogram HOG3D, motion boundary histogram MBH, and extended accelerated robustness Descriptors such as feature ESURF are used in combination. [0004] The extraction of local features mainly consists ...

AI Technical Summary

Problems solved by technology

Although this method takes into account the relationship between different features, it only considers the static features and does not consider the dynamic features of the behavior subject, and cannot accurately describe the movement of the behavior, resulting in low accuracy of behavior recognition results.

Images