Gait Recognition Using Hide Markov Model

Gait recognition refers to automatic identification of an individual based on his/her style of walking, and it is a new biometrics recognition technology. Prominent characteristic of gait recognition is identification of an individual from a distance, so it has a broad application prospect at security system, human ID management, and digital surveillance.Hides Markov model (HMM) is a model based on statistics, and it has the strong mathematics foundation. HMM has a well performance at forecast and simulation of random data sequence. Using HMM to the gait recognition can get the dynamic information from a stance to another stance, and obtain a well recognition effect. In this thesis gait recognition processing includes three central parts: preprocessing of gait sequences, feature extraction and HMM modeling.The preprocessing of gait sequences is extract human motion from gait video. It mainly includes motion detection and morphological postprocessing. An improved approach of background subtraction is proposed, and then improves quality of foreground image by morphological process.In the feature extraction stage, recognition system calculates the outer contour of a human silhouette feature using area feature of silhouette, then obtains five exemplars from a gait cycle by K-means algorithm and compresses feature vectors by Euclidean distance between exemplars and each frame.In the recognition part, a HMM with mixture of Gaussians outputs is constructed for each individual in a enrolled gait database. First, we initialize a Gaussian mix model for training image sequence with K-means algorithm, then train the HMM Parameters using a Baum-Welch algorithm. We calculate the output probability of the unknown gait sequence to all enrolled HMM and choose the person corresponding to maximum output probability as its classification identification.The proposed approach method analyses and captures human gait structure and the dynamic information. Experiments are implemented on gait database of CASIA and the experimental results show that the proposed approach is quit robust and has a high recognition rate in CASIA database.