Study On Object Silhouette Extraction And Feature Dimensionality Reduction In Gait Recognition

Gait recognition is a technology of biometric recognition, which recognizes people by the way they walk. Compared with other technologies of biometric recognition, gait has its own unique advantages, such as recognizing people without contact. The gait feature is not easy to disguise and it can be perceived from far away. With the increasing demand for intelligent visual surveillance and monitoring systems in security-sensitive environment, gait recognition has attracted a wide range of research interests.Gait recognition is mainly composed of gait silhouette extraction, feature extraction and classification. This paper mainly includes the following aspects:①Based on the characteristics of gait sequences, a gait silhouette extraction algorithm based on the probability distribution of the intensity value of background pixel is proposed. The gait silhouettes of each sequence change over time. The method constructs background pixel distribution models for each pixel with respect to time, and uses hypothesis testing to segment moving objects. The experiment results demonstrate that the proposed algorithm not only can extract human silhouette completely and accurately, but also can remove noise and shadow and improve recognition rate. The algorithm can be directly used on RGB space or grey space, without changing color space or building the background model, it is computationally simple to implement and has high real-time processing performance.②As for the gait feature extraction, the outer contour of the silhouette is projected along the width of the body to obtain width projection vectors which are then translated into one dimensional vector as the gait feature. The data size is big and high-dimensional, so it is hard to deal with the identification if the width vectors are used directly as input. The Principal Component Analysis (PCA) algorithm and Supervised Locally Linear Embedding (SLLE) algorithm are used to reduce the dimension of feature vectors from 201 to 14 and 19 respectively. The two dimensionality reduction methods are complementary, the internal structure of the high-dimensional gait characteristics is maintained in low-dimensional space, and the operating efficiency of the algorithm is improved.③Gait recognition is performed by the k-NN classifier, and the Bayesian combination rules are used to fuse the feature information extracted by the Principal Component Analysis (PCA) algorithm and Supervised Locally Linear Embedding (SLLE) algorithm. The experiments demonstrate that the proposed algorithm performs better than one classifier system in terms of classification and verification capability.