Human Behavior Understanding Based On Walking Trajectory In Home Intelligent Space

Behavior understanding is a prospective research topic, which involves a number of disciplines such as sensor technology, image processing, pattern recognition, artificial intelligence, etc. In the upsurge of smart home and intelligent wearable device, behavior understanding is not only where the "smart" stands, but also one of the key technologies that urgently need to be broken through. In this paper, human walking trajectory is the breakthrough, and a set of feasible schemes to understand human behavior are proposed. The main research works are organized as follows.To solve the problem of acquiring human walking trajectories, a vision based method is proposed, where human state is divided into "walking" and "static" ones. In "walking" state, human is detected using mixed Gauss background model; while in "static" state, human search-box is kept using particle filter and meanshift algorithm. After stable human tracking, human location is obtained by using binocular vision. With the intelligent space as experimental environment, the method is verified. With this method, lots of data is collected that lays the foundation of further research.To solve the problem that the behavior level can-not be divided independently, a hierarchical trajectory segmentation strategy combining sliding window segmentation and key point segmentation is proposed. The strategy can automatically divide the trajectory among two behavior levels, i.e. "action" level and "activity" level. The "action" level focuses on human’s instantaneous motion state; therefore, "sliding window" is used to extract the trajectory belonging to a motion. The "activity" level focuses on human’s complex activities; therefore, "key point" is used to extract the trajectory belonging to a specific activity. According to the trajectory feature, the segmentation method can be selected automatically, so that the behavior level can be divided independently.Action level:To meet the requirement of on-line trajectory trend forecasting, an online trajectory analysis scheme based on Hidden Markov Model (HMM) is proposed. The "complete trajectories" are used to train the model, and the "window trajectories" are used to evaluate the model. In order to achieve the matching between "window trajectories" and "complete trajectories", the initial state probability matrix of HMM is set as the uniform distribution. In this way, the state transition probabilities and observation probabilities can play decisive roles, and the effect of the initial state is ignored. Morever, instead of building one HMM for a "A-to-B" path, or building another HMM for a "B-to-A" path, a special measuring method is taken, which makes model "A-to-B" able to point out both "A-to-B" and "B-to-A" trajectories from others. The learning time and evaluating time for HMM are reduced as well.Activity level:To meet the requirement of describing and recognizing complex behavior, the "two-Dimension Grid Trajectory Histogram (2D-GTH)" is proposed. 2D-GTH is built on the basis of analysis results from "Action" level. It refines the trajectory features and has a simple structure. Accordingly, the "Grid-Trajectory Probabilistic RoadMap (GTPRM)", which is a tool to understand 2D-GTH, is proposed. It shows the probability statistical model of grid trajectory in the form of a map. Finally, in different evaluation systems of the histogram, the effectiveness for complex action recognition based on 2D-GTH is verified.Based on the above works, relevant software is developed. After quite a few experiments, it shows that the proposed behavior understanding scheme based on walking trajectory has excellent effect on both real-time forecasting and complex behavior analyzing.