Hidden Markov Model-Based Approach for Recognizing Human Behavior and Application in Robotic Home Care

• Main Authors: Wen-che Yu, 于文哲
• Other Authors: Ming-jong Tsai
• Format: Others
• Language: en_US
• Published: 2014
• Online Access: http://ndltd.ncl.edu.tw/handle/2fswg5

碩士 === 國立臺灣科技大學 === 自動化及控制研究所 === 102 === Recent years, how to infer human behavior from video automatically has become an important topic in many advance video surveillance applications. This study applies a HMM (Hidden Markov model)-based methodology to detect some specific-behaviors of a person in a domestic environment. In this study, a behavior is composed of a series of the static image-frames. For each frame, a Convexity Structure is used to enclose a human body’s shape after segmenting the blob. With the Convexity Structure, this study clearly defines three feature parameters, Width Height ratio, Height ratio, and Convex Hull Perimeter ratio as a three-dimensional vector. Firstly, more than 4000 blob types are analyzed from the possible static postures captured from each interested behavior. Then, they are clustered into 6 feature-postures by using K-mean method, and a codebook is created for mapping the image blobs into one of 6 feature blob-types during training/recognition stage. Consequently, the time-sequential blobs are converted to a feature-vector sequence and transformed into a symbolic-sequence by the codebook. Thus a learned HMM can be obtained by this symbolic-sequence from a given specific-behavior. After that, the learned HMM is used to detect the unknown behavior to investigate which behavior it is. A system is implemented to automatically recognize five different behaviors in domestic environment (Walk, Walk Fast, Bend, Pick up an Object, Jump) and one abnormal behavior (Fall Down). The system is trained by using 120 known video clips which record the behaviors of 1.8m-height and 1.4m-height human. Subsequently, another 240 video clips are used to verify this system. This system also includes a humanoid robot to simulate home care situation. According to a particular human behavior, the robot can perform a correspondent reaction. From the experimental results, an overall recognition rate of 95% is achieved, and the system can detect abnormal behavior accurately.