Research On Human Identification Based On Gait Analysis

As an attractive direction in biometric and computer vision in recent years, gait recognition aims to recognize individuals by the way they walk. Comparing to other biometric features, gait is non-contact, non-invasive, easily acquired. Furthermore, gait is difficult to hide and to be disguised. With these advantages, gait recognition has a good foreground in many domains, such as visual surveillance and medical diagnosis, so it has received an increasing interest from researchers all over the world.A gait recognition system consists of four parts: preprocessing of gait sequences, period detection, feature abstraction and recognition. Feature abstraction is the most important thing on which our study focuses. Working on gait recognition methods, this thesis concentrates on the following topics:The status of art of gait recognition is introduced; passive actors and performance evaluation are generalized as following. After that, our focus is concentrated on the main points of the four parts of gait recognition. Considering kinds of motion detection methods, the background subtraction is used in motion detection. Morphologic operation and scale transformation are used to standardize and centralize the gait sequences. A new period detection algorithm is proposed, and experiment demonstrates it works efficiently and exactly. By using the period sequences, a gait energy image (GEI) is created to depict gait.Feature abstraction of gait is researched, and approaches can be broadly divided into two classes. One is model-free method and the other is model-based method. Model-free method does not require prior model, just assumes some implicit notion of what is being observed and characterizes gait by the statistics of the spatiotemporal patterns generated by the walking person in the image. Model-based method assumes a priori models, and matches the 2D image sequences to the model. Problems of using LDA are analyzed, and an improved LDA method is adopted to search a better feature description in zero space of gait feature. After that, we get a gait feature with low dimension.Each sample in the probe is matched to samples in the gallery by using KNN classifier and Euclidean distance measurement, and a gait based human recognition system is achieved. Experiment on gait database of CASIA (including 124 samples) shows that this system achieved a recognition rate of 98.7% on side view.Based on this thesis, some expectation and tentative are proposed as follows: fusion of gait feature in multiple views, 3D modeling & usage of multiple cameras, data fusion and performance evaluation under complex background with big gallery.