Human Action Recognition In Video Surveillance

Human action recognition technology has a wide range of application prospects in security monitoring,virtual reality and other fields.In recent years,it has become one of the main research directions in the field of computer vision.At present,the mainstream action recognition algorithms are mainly divided into three steps: motion region detection,feature extraction,and action recognition.In order to monitor the abnormal operation of humans in the computer room monitoring video,these three key technologies and human motion recognition methods in indoor surveillance videos are studied.Action recognition method based on dense interest point optical flow trajectory and optical flow constrained AutoEncoder has been studied in depth,the specific contents are as follows:(1)A motion region detection algorithm based on background subtraction is proposed for the detection of motion regions in video.This algorithm firstly uses the improved Vibe algorithm to detect moving objects,ghost phenomenon is eliminated by the inter-frame difference method,and using the Otsu to dynamically determine the segmentation threshold of the foreground and background.Secondly,the number of foreground pixels is counted to filter out redundant frames that do not have human activities.Finally,a weighted summation of the foreground images is added in a time sequence to obtain a motion intensity accumulation image,and the circumscribed rectangle contour of the motion intensity accumulation image is taken as a motion region of the human body.The manually labeled motion region is compared with the detected motion region,experiments show that this method can effectively extract the motion region,and the improved Vibe algorithm has better detection accuracy.(2)For the feature extraction problem,a feature extraction method based on dense interest point optical flow trajectory is introduced.Firstly,this method combines feature point detection algorithm with dense sampling method,and proposes a dense interest point detection algorithm.Secondly,dense feature points are traced by calculating dense optical flow fields,and the optical flow trajectory of dense interest points over a period of time is obtained.Finally,referring to the traditional space-time feature descriptors,the spatio-temporal pipeline based on the dense optical flow trajectory is divided into several spatio-temporal grids.Local descriptors such as trajectory descriptors,HOF,HOG,and MBH are calculated within the spatio-temporal grid.These feature descriptors are series connected to obtain an action feature descriptor based on a dense optical flow trajectory.Experiments show that dense optical flow trajectories have better ability to describe motion than sparse motion trajectories(3)A feature extraction algorithm based on optical flow constrained AutoEncoder is proposed to solve the problem of feature extraction.AutoEncoder are widely used in feature learning algorithms,however,the features learned from the AutoEncoder have the disadvantage of not distinguishing between static information and motion information.The algorithm uses optical flow information as a regularization term of the AutoEncoder,and a regularized AutoEncoder based on the optical flow constraint is designed.The AutoEncoder network takes video pixel information and optical flow information as the input of the network,and calculates the network parameters through back-propagation algorithm and gradient descent method.Experiments show that the features merged with optical flow information not only have good sparsity,but also have better dynamic characteristics.(4)Aiming at the problem of action recognition,BOF model is used to model the action features,and nonlinear SVM is used to conduct action classification.The two feature extraction algorithms are tested on the public database KTH and Hollywood2,Analyze the influence of various parameters on the recognition result.Finally,the two feature extraction algorithms are compared on the simulated database,the feature recognition rate of the optical flow constrained self-encoder is 86.4%,and the feature recognition rate based on the dense optical flow trajectory is 84.3%.Experiments prove the effectiveness of the two feature extraction methods,the optical flow constrained self-encoder feature is superior in performance to the dense optical flow trajectory feature.