Motion Capture Data Based Human Behavior Segmentation And Annotation

With the increasing demand of realistic human motion data, masses of MoCap databases have been springing up in the last few years. These MoCap data can be utilized in the many fields, like production of video games and animation movies, physical rehabilitation and ergonomic analysis. In order to effectively and efficiently take advantage of MoCap data in these real scenarios, long motion sequences have to be segmented precisely into distinct behaviors and then be annotated with the right behavioral semantics.Since manual segmentation and annotation is prohibitively tedious and time-consuming, automated methods are indispensable. But several issues contribute to the challenge of automatic segmentation and annotation of human motion, such as the incalculable variation in behavioral periodicities, the intricate representation of articulated body postures, and the countless combinations of all possible behaviors. In this thesis, a novel method based on Posture Histogram in Sliding Window is proposed to solve behavior segmentation and a new method based on Behavior Template is presented to work out behavior annotation.In both the problems of behavior segmentation and behavior annotation, it is necessary to find a way to characterize behavior features of motion sequences. Therefore, this thesis first defines three sets of new posture features and constructs a new compact representation of behavior features, Posture Histogram, by calculating proportions of motion frames in subintervals of posture features. Moreover, a selective combination of posture features is proposed for the first time to derive the Posture Histogram.To locate segment frames by analyzing sudden changes of behavior features, sliding window is applied to extract subsequences. And in order to obtain conspicuous and stable behavioral characteristics of the extracted subsequences, a novel way is put forward to tune sliding window by analyzing steady states of human behaviors. Benefiting from the great clustering property of posture histograms of extracted subsequences, behavior segmentation problem is tactfully simplified to the detection of outlier subsequence whose central frame is taken as the segmenting point. And the Local Outlier Factor (LOF) algorithm is first adopted to pick up the outlier subsequences. The results of experiments show that the novel behavior segmentation method can achieve 98% precision rate and 97% recall rate.In behavior annotation, Behavior Template is proposed as the representative of behavioral semantics. Considering different body parts have different levels of contribution in a behavior, a selective posture histogram is built from a set of most informative posture features, which have high variances. Then, Behavior Template is created on the selective posture histogram by screening out the parts of dimensions corresponding to the posture feature subintervals with high variances and low mean values. Finally, Jensen-Shannon divergence is utilized to measure the similarity between the posture histogram of a motion sequence and Motion Templates. And the motion sequence is annotated with the behavioral semantics of the Behavior Template, which has lowest Jensen-Shannon divergence. The results of behavioral semantics annotation experiments indicate that the Behavior Template based method can achieve 97% precision rate, which is 43% higher than that of the original posture histogram.