Article Preview
Top1. Introduction
In recent years, gait has become a research hotspot in the field of image processing and recognition as a biological feature. Gait recognition is based on the posture of people walking to identify the classification. Gait has the characteristics of non-invasive, unique, easily to collect features, so it can be used for identification (Xiao, Liu, & Dongxu, 2012). Therefore, gait recognition has attracted more and more attention, and has wide application prospects in video tracking, medical care, and so on.
The methods of gait recognition are broadly divided into two types: model-based algorithm and non-model-based algorithm (Zhang, Jiang, Wei, Zhang, & Shi-Jie, 2012; Liu, Jia, & Zhu, 2009). At present, non-model-based methods are more popular, which common are the shape feature, time and space characteristics (Chen, Shi, & Chun-Hai, 2012; Yang & Liu, 2012). Most scholars believe that the shape feature is one of the most effective methods for gait recognition. In addition, the gait image of each frame has different shape features in the process of walking. Therefore, it is very important to describe the shape of the gait image and extract the features of each frame. The common shape feature descriptions are shape signature, invariant moment, Fourier descriptor, and so on. In the (Yu, Wang, Hu, & Tan, 2004), the Fourier descriptor of the key frame is used to express the gait characteristics. In the (Ben, Xu, & Wang, 2012), the Hough transform and the Trace transform are used to obtain the shape features of the gait. In the (Tang, Song, & Jiang, 2008), the static information is obtained by invariant moments, and it can obtain a good recognition rate when combined with the dynamic information. However, these methods ignore the local details of the shape and anti-noise performance is poor. In the (Chen, Ma, Huang, & Gao, 2007), the shape context descriptor is proposed to represent the human body contour feature, considering the pixel position vector relations, achieving good recognition rate, and with the increase of the number of sampling points, the recognition rate is improved in a certain range. However, this method only considers the contour points, and does not reflect the joints of the lower limbs with abundant information. With the increase of the sampling points, the calculation becomes larger.
In this paper, based on the shape of the context, combined with the model-based approach, select the contour points and lower limb joints as sampling point, which can better articulate the relationship between the joints and contour points; in order to reduce the computational complexity, do not select the shape context matrix of each point as the feature, but in the polar coordinates, with different sampling point as the origin, the number of statistical sampling points in each sector is counted as the feature after traversing all the sampling points. The final shape context feature is no longer a matrix, but a vector. When the polar coordinates are determined, the dimension of the vector is determined and does not increase as the sampling point increases.
Finally, the computational complexity and recognition rate of the original shape context are improved.