Research On The Key Technologies Of Object Tracking And Facial Expression Recognition

The thesis focuses on the object tracking and the facial expression recognitionresearch in computer vision. In the object tracking under simple conditions, two basicframeworks of video events detection are established based on the motion or the changeof the targets; under complex conditions, the research is focused on mean-shift andparticle filter tracking framework in two major components: multi-feature fusing andobject model update, several new effective methods have been proposed whose aims areto deal with the difficult problems in object tracking. For the expression classificationproblem, we focus on the key algorithms for feature extraction. We introduce HOG(Histogram of Oriented Gradients) feature and LBP (Local Binary Patterns) feature tothe expression recognition that are widely used in images classification recently years.Besides, this paper is in-depth study of the smile facial expression classification,features fusion, etc. Generally, the main contributions of the thesis are as follows.1. In the moving targets detection and tracking，two novel frameworks areproposed to process the video for events detection：(1) A region based abnormalbehaviors detection approach of the human body is proposed by simple informations ofthe targets such as the minimum bounding rectangle and center mass of object. Thenormal and abnormal features databases are established throught K-means clusteralgorithm, while the abnormal behaviors in video are detected by using the changedinformation of the targets.(2) A motion based detection approach to count the numberof pedestrians crossing in the gungway scene is proposed by using the improvedcentroid algorithm for multiple objects tracking． With the study on movingtrajectories，the approach can get the regions and the moving direction of the movingobject，then count the number of people．These two approaches need no precisedetection and tracking of object, so they can satisfy the on-1ine video-processrequirement．2. Object tracking algorithom results in a poor performance in complex scenes. Tosolve this problem, an object tracking method based on multi-features fusion ispresented. For the disadvantage of the color distributions that it omits spatialinformation, while the text feature describes the spatial distributions of local region ingray, they are complementry. So the algorithm builds the object model that includes thecolor distributions and the texture features extracted by Local Binary Pattern (LBP).Meanwhile, the robustness of the tracking is strengthened by integrating the mean-shiftinto particle filter, and performs a feature fusion in both of them with two common used fusion rules are employed respectively, thus overcoming the degeneracy problem andresulting in low computational cost.3. To improve the object tracking performance under complex scenes such asillumination changes, target appearance changes and when occlusion occurs, theresearch focuses on the two major components of tracking algorithoms:(1) Accordingto the shortages of multiple features fusion with fixed-weighting policy, an algorithmfor fusing multiple features adaptively under particle filter tracking framework isproposed. The tracked object is represented by the fusion of all features under linearweighting, and a new method to estimate the fusion coefficient is also proposedaccording to the weight distribution of all particles as well as their spatial concentrations,thus improve the reliability of multi features fusion.(2)To overcome the shortage oftraditional template update strategy, a dynamic template updating strategy is designedfor multiple features template. According to the degeneracy of each feature，this strategyis used to adjust the update speed of each feature template adaptively. Besides, featurewith high confidence is used to detect occlusions thus decreasing the influence of partialocclusion and appearance changes, partially avoids the model drift caused by updateprocedure.4. The main technologies of facial expression feature extraction and classificationare generally described. Considering the three deficiencies of local binary pattern (LBP),the approved LBP is adopted at the stage of feature extraction, which reduces the featuredimensions and experiment results show that the method has a fast speed and goodability to classify the face expressions. After an in-depth study of themulti-classification SVM, the expression classification method based ontwo-against-two SVM is proposed for the defects of the traditional classificationmethods. Based on the research of four-category classification method, thetwo-against-two classification method realizes fast classification of six basic facialexpressions with a relatively fast speed and a high performance, reducing theaccumulation of classification error.5. To the question of smile expression recognition，a novel method base on featurefusion is studied. A comparative study of LBP features and HOG features was presented.Thus the fusing of the two features for smile expression classification was conducted.The research about the fusion algorithms based on series connection, canonicalcorrelation analysis (CCA), and the discriminative CCA (DCCA) are discussed in detail.The experimental results demonstrate that the fusion features can use the complementary between texture features and shape features of facial expression, thussignificantly improve the effectiveness and superiority of smile expressionclassification.