Feature Extraction of Human Body for Gait Analysis and Recognition

• Main Author: 曹瀠方
• Other Authors: 邱瀞德
• Format: Others
• Language: zh-TW
• Published: 2013
• Online Access: http://ndltd.ncl.edu.tw/handle/09347031000344713826

碩士 === 國立清華大學 === 資訊工程學系 === 101 === Human walking behaviour can be affected by the disorders, physiological condition such as the chronic obstructive pulmonary disease(COPD) patient walking with large breathing, the extremely fatigue people walking in static pace rhyme. The walking behaviour can not only be used in gait analysis to detect the patient condition, but also be used in gait recognition. Everyone has his own characteristics and habits under the comfortable walking status such as the pace velocity, the waving of arms, the lifting of legs, the humpbacked and else. We base on the geometry of human to segment the silhouette on image processing in order to construct the model of body parts such as head, upper torso, lower torso, arms, and legs. We utilize the mass of body parts to develop our proposed features of angles, ratios, and pace. In gait analysis, we design different experiments for the people with respiratory obstruction, with bending and in normal. When people with respiratory obstruction, they have large breathing because of lacking oxygen. Also, the angle of head, and upper torso has the $2.49$ times larger than normal. In humpbacked condition, the angle is $5.5$ times as large as the normal condition. Other fatigue detection experiment, we observe that people in extremely fatigue condition has the pace rhyme lower than $3$ because the pace becomes static and slow. It is because that people have tired physiological condition and they can not walk like normal walking with energy. As a result, people walk with the lowest power manner to reduce their energy consuming. In the velocity experiment, we request the people to walk fast and slow, and the velocity of fast case is $1.347$ times as large as the slow normal walking condition. Taking these above experiments into consideration, we can discriminate some symptoms of patient, the poor physiological condition. When the detected features are larger or lower than the threshold which depends on the normal condition, we can send a notice. We not only utilize our features on gait analysis, but also on human gait recognition. We experiment on the two benchmark databases which both have the walking images above 100 different people. Our recognition rate on normal condition can both achieve $90\%$ under same testing and training database as a reference. Although, our recognition rate is not better than model-free approaches, it can display that our proposed method can discriminative the people under different environments when walking. The experiment comparing with model-based method under normal walking status has highest performance with $96.97\%$ accuracy using CASIA database B, and we could use it as a reference for gait recognition rate to compare with other model-based methods.