Gait Recognition Based On Support Vector Machine

Biometric recognition is an identification technique using physiology behavior or behavioural characteristics. Current individual human recognition methods such as face, fingerprints, or iris biometric modalities, generally require a cooperative subject and physical contact or close proximity. These methods cannot reliably recognize noncooperating individuals at a distance in the real world under changing environmental conditions. But gait, which concerns recognizing individuals by the way they walk, is a relatively new biometric disregard these disadvantages. Gait is the manner an individual walks. It's a complex behaviour, which provides plenteous information for identifying individuals.Traditional classifiers such as Bayes classifier and nerve network classifier, donot perform perfectly when samples are not enough. They can work well if samples close to infinite. For gait database, the amount of samples is very finite and the methods above are not applicable. The theory of Support Vector Machine (SVM) is a statistical theory bases on small samples. It can solve the problem of small samples learning better and be more ubiquitous.In this paper, we try to apply SVM to recognize individuals by gait. Gait identification bases on features acquired from walking process and the features cannot acquire from one single image of gait sequence. It brings the problem of"dimension disaster"and serious operation cost. Gait Energy Images (GEI), which is a method to add and average a period gait images into a single image and contains the gait information about frequency, phase, figure and so on, can decrease data and save gait information better simultaneously. Principle Component Analysis (PCA), Kernel Principle Component Analysis (KPCA), and Fisher Linear Discriminant Analysis (FLD) are performed on Gait Energy Image to decrease data ulteriorly.The method was evaluated on Carnegie Mellon University (CMU) dataset and University of South Florida (USF) Challenge gait dataset. Results show that the method has encouraging performance. CMU dataset has much better indentification rate because the images are more legible than USF's.