應用步態分析於身份鑑別之研究

Abstract

本論文提出一個自動的身份識別系統，主要目的在於減少人力成本的耗損，並提高監控系統的安全性與可靠性。不同於其他的生物識別系統，例如人臉辨識、指紋辨識、虹膜辨識等等，利用姿態進行身份識別不需與使用者互動、不需近距離拍攝、也不需要太精密的攝影器材即可進行分析，因此步態分析一直是視覺監控領域的熱門研究主題。但由於步態辨識遠距離即可完成，因此相對其他生物識別系統來說受到外在因素的影響也較大，例如：戴帽子、穿著不同等等，所以步態辨識仍然是一個很受挑戰的領域。 在此，本論文的貢獻有兩點：第一，本論文結合多種方法，使用多面向與多角度來分析姿態，不僅平均抽取人體各個部位的特徵，更進一步將特徵分為靜態特徵與動態特徵，以確保本系統面對衣著變因的影響時，還能保持一定的辨識率；第二，本論文提出兩階段式的身份鑑別系統。第一階段主要是利用外輪廓、手部擺動和腳掌離地角度三個特徵來進行候選人篩選，而篩選的目的在於減少候選人，以提高整體辨識率；第二階段則是將輸入的步態影像與第一階段篩選出來的候選人進行匹配，並針對每種特徵不同的重要性對這些候選人進行不同的權重給分，分數最高的即表示與輸入的人物最為相似。從實驗結果來看，本系統確實有不錯的身份識別結果。

This thesis is a study on individual recognition using gait analysis. In order to reduce the human resource cost on one hand and improve the security and reliability of the surveillance system on the other, we propose an automated individual identification system which can not only recognize identities of individuals but also detect strangers automatically. Different from other biometric systems, such as human face detection, fingerprint identification, iris recognition, etc., the proposed system uses gait to recognize identities of individuals. In addition, the proposed system needs no interaction with people, nor close sensing distance, nor fancy equipment. Because of these advantages, gait analysis has always been a popular research topic. Furthermore, there are two contributions in this thesis. First, features in this thesis are extracted averagely from head to toe to alleviate the factor of clothing which would cause the distortion of appearance on an individual. Second, we use a two-stage method to recognize identities of individuals. The first-stage is designed to reduce candidates. In the second-stage, we match the individual extracted from the video with candidates selected from the first-stage, and assign weighted scores to the candidates based on their similarities. The candidate who gets the highest score is the most similar one to the individual extracted from the video. According to the results of experiments, the method proposed in this thesis has a good performance.