Constrained Sparse Coding Methods For Human Action Recognition In Video

Human behavior recognition technology, as an important branch in computer visionfield, is an emerging technology mainly focusing on analyzing and recognizing thehuman behavior in video. Due to its widely application prospects in intelligentsurveillance, vitual reality, auxiliary training, human-computer interaction etc, it hasattracted widespread concern of scholars over the world recent years.Sparse coding, one of the new theory developed along with the compressed sensingtheory, has made tremendous achievements in the field of image denoise, imagesegmentation, image classification and face recognition and has been demonstrated tohave excellent performance in dealing with computer vision problem.Due to the complexility of human motion and the variability of the background invideo, it made the analysis and recognition of the human behavior in video became aextremely challenging task. Althoug human behavior recognition attracted more andmore attention from scholars and has made some significant achievements recently, ingeneral, it still at the basic research stage and still far from the pratical application. Mostof the sholars now focus on the research of simple behavior recognition under simplebackground,we call it action recognition. Rencently, due to its good performance invariable viewpoint, complex background and dynamic light condition, local space-timeinterest points based recogniton method has became the most popular method recently.Focusing on the action recognition problem in video, this article carried outresearch on human action recognition using sparse coding theory based on localspace-time interest points and mainly made two contributions to the research andapplication of sparse coding theory in human action recognition in video. Thecontributions are listed as follows:(1) Taking into account the fact that the spatial locality distribution of the featuresin video present a certain geometry structure, we proposed a method called SLCLC(Spatial Locality Constrained Linear Coding) based on the LLC. SLCLC method takefeatures’ locality in human body ROI into account when coding the features and makethe coding result of the features in same part of body similar while the coding result ofthe different part of the body unsimilar.As a result, this coding method has the ability todescribe the semantic relation between features. Experiment shows that this methodenhance the recognition accuracy obviously in video when its body ROI is regular insemantic aspect.(2)Taking into account the fact that the video frame sequences are locally similarity,inspired by LLE (Local Linear Embedded) method, we proposed another coding methodcall LSTCLC (Local Spatial Temporal Constrained Linear Coding) based on LLC.LSTCLC limit the coding domain of features in the local neighborhood and take the spatial-temporal distance and cocurrence between features and dictionary basis intoaccount when coding those features, which enhance the spatial-temporal relationship ofcoding result between features. And as to the details of implementation,we compute thespatial-temporal distance and co-occurrence probability between features and dictionarybasises by using3D gaussian kernel function and by constructing co-occurrencerelationship matrix between each class of features respectively. Experiments show thatthis method is more adaptable in complex action scenes.