Human Action Recognition Based On Random Fields

Human action recognition is a hot topic in the field of Computer Vision and PatternRecognition. It extracts the moving body from the video sequences w hich include human actionsby object detection and tracking, and on these bases it constructs a classifier to model any actionaccording to the features used to represent the visual characteristics of an action, so it can realizethe auto-recognition of an unknown action sequences. Human action recognition has a goodprospect of application in many fields, such as video surveillance system, digital entertainment,human-machine interaction and so on.In this paper, we did some research work about human action recognition method, the maincontents included the following aspects:1)object detection and tracking: the moving human bodywas extracted from the background by modeling the background of the video sequences andapplying the background subtraction method. Based on this, we adopted the efficient Mean-shiftmethod to track the human body in action video sequences. Until now, the consecutivespatiotemporal region of human body has been extracted, it would be a basic part of human actionrecognition.2) extraction of image features: The main visual features studied in this paperincluded Gist feature, Shape context feature and BoW feature, these features were computedrespectively for all frames to represent the visual characteristics of human actions.3) modeling theactions: in this part, the construction, learning and deducing of three kinds of randomfields(conditional random fields, hidden conditional random fields, latent-dynamic conditionalrandom fields) which could model the sub-structure of time sequence were introduced, andrealized the recognition of human actions based on these models, then the performances of thesealgorithms were analyzed.In this paper, several experiments of different feature-model combinations were designed.These experiments were run with Matlab. The performances of the above features and modelswere tested on two common and public dataset (Weizmann dataset and KTH dataset). Theexperimental results showed that the features adopted in the experiment could represent humanstructure and visual characters well, and random fields were good for various human actionsrecognition for that they could model the sub-structure of the action sequence. Furtherexperiments and analysis demonstrated that HCRF were more flexible in modeling non-structureaction sequence as the introduction of hidden states, so that its recognition performance of timesequence was much better compared with CRF and LDCRF.