A Research Of Behavior Recognition Algorithms Based On Mult-features Fusion

Human behavior recogniton is receiving increasing attention from computer visionresearchers. The computer vision includes three aspects, which are human objectrecognition, object tracking and behavior recognition. Among them, Human behaviorrecognition is the highest-level part. Robust behavior recognition algorithm hasimportant theoretical significance and great application value, which can be widely usedin the field of intelligent video surveillance, video retrieval and human-computerinteraction. At the same time, the difficulty of the research is mainly reflected in thecomplex background, multiple types of actions, large amount of data and real-timerequirements.Behavior recognition process includes three important aspects, which are thedetection and tracking of the human body, feature extraction and behavior recognition.In recent years, behavior recognition technology based on the spatio-temporal interestpoints is widely applied，which has the advantage that it does not depend on theunderlying human detection and tracking algorithm, but depends on the aspects offeature extraction and behavior recognition. This thesis follows the behavior recognitionmethod based on the spatial-temporal interest points. As to extracting feature, first, wedetect spatio-temporal interest points in the three-dimensional video samples, andintroduce the background difference method to obtain the region of interest. Secondly,the region of interest and the minimum region of STIPs are fused to obtain the finalregion of interest. Once again, we extract the local feature information, such ashistogram of oriented gradient, histogram of oriented optical flow and wavelet energyinformation. Finally, this thesis uses the cumulative histogram to merge the sequenceimage features and generate video eigenvectors. As to behavior recognition, this thesisadopts the direct classification methods of KNN and AdaBoost. KNN algorithm issimple, but low recognition rate. AdaBoost algorithm has a high recognition rate, but ismore complicated because it needs a weak classifier training process.Compared to previous work, the contribution of this thesis is mainly reflected inthe following aspects: 1. Proposed a new feature representation called STIG（Spatial-Temporal InterestGrid). STIG is based on ROI and STIP. Through combining global features and localfeatures, it can effectively hold the space connection information among eachspatio-temporal interest points. As to ROI, it is the fusion of the global foregroundmotion region and local minimum rectangle of STIP region, which obtained by usingthe bidirectional analysis method. In addition, the ROI provides the basis for otherfeature extraction.2. Proposed a method that combining PHOG, PHOF and PHOW feature todescribe the behavior. The PHOG is a shape feature representation method. And PHOFcan describe the motion features efficiently. In addition, this thesis introduces a newdescriptor named PHOW, which is able to get the energy field distribution in a variousfields of an image. For the natural background, PHOW has strong feature presentationcapabilities. Therefore, for the shape and appearance description, PHOW and PHOGcomplement each other.3. Introduced cumulative histogram to integrate the feature information insequences and to form feature vectors. The joint feature maintains spatial information,and the cumulative histogram maintains the time information.In the experiment, this thesis uses KTH and Weizmann data sets as benchmarkdata sets, and a natural background data set UCF as a challenging date set. KNN andAdaBoost are used as the classifiers. For the new feature descriptor we proposed, wediscussed various quota using experimental data, and verified the rationality andefficiency of the proposed algorithm.