In this research, we apply artificial intelligence and statistical techniques towards observation of people, leading to modeling of their actions, and understanding of their expressions and intentions. We propose to develop methodologies to understand human behaviors intelligently by learning from demonstration. The techniques developed will be incorporated into practical systems in application areas including learning of human emotional expressions, classification of pedestrian trajectories for surveillance, recognition of human sport and fighting actions, and network architecture for distributed recognition modules.; First, we have developed a system that can automatically estimate the intensity of facial expressions in real-time. Based on isometric feature mapping, the intensity of expressions can be extracted from training facial transition sequences. Then, intelligent learning algorithms including cascade neural networks (CNN) and support vector machines (SVM) are applied to model the relationship between trajectories of facial feature points and expression intensity level.; Second, we have developed an intelligent surveillance system that can automatically detect abnormal pedestrian walking trajectories in real-time by learning from demonstration. By using support vector classification, we can identify the trajectory points at which the observed pedestrian is performing abnormal walking motions. By utilizing a stochastic similarity measure based on hidden Markov model (HMM, the normality of the shape of the entire trajectory can be determined. The outputs of both learning mechanisms are combined by a rule-based module to arrive at a more reasonable and robust conclusion.; Third, we have developed a tracking and learning system that is capable of classifying full-body actions that occur in sport videos and detecting the actions of person-on-person violence. A tracker is developed to locate the positions of human head and hands by using background subtraction and silhouette analysis. The motion data is then compressed by using principal component analysis and independent component analysis. The motions performed by the people in the scene can be recognized using support vector classification.; In terms of networked human motion understanding, we have developed a service-based architecture to enable the flexible and reconfigurable connection between the interacting components in distributed networks. The proposed network structure can be used to support distributed analysis of human motion and intention.
展开▼
