Multi-perspective Gait Recognition Independend Of Clothing And Belongings

In recent years, information technology is accelerating the schedule of making information security and public safety increasingly highlighted its importance, especially for the troops on the occasion which requires higher safety, how to accurately identify a person's identity, to protect information security has become a key social issue that must be addressed. Biometric technology is more and more favored by people of all ages because its simple, fast, identity finds reliable and accurate, ease to integrate with security, monitoring, management and many other advantages. In this backdrop, gait recognition is attracted attention.Compared with the traditional biometric features, gait recognition has the following three advantages: The first, Non-contact, it no need complex by person who is to be distinguished and can complete the acquisition in the case not to be found; The second, long-range nature, without close contact to obtain and recognition; The third, authenticity, gait is difficult to conceal. At present, most gait recognition algorithms are based on the front or side perspective and not to be obscured by clothing or belongings, this makes extensive use of gait recognition has been greatly restricted. To combine the work of Chinese People's Armed Police Force(CAPF) at the sudden terror, guard on duty, assist the public security organs to arrest and so on, this article is intended to find a universal gait recognition algorithm that is not intended by perspective, clothing, belongings and so on, enables gait recognition to be more widely used. The gait recognition can be applied to the daily work of CAPF. In this article, the main work done and the results achieved:(1) Gait preprocessing. In this section, gait detection?image normalization and to determine gait cycle are reseached. Contrary to background subtraction that for gait detection innovation, the moving human images obtained by conventional background subtraction method is prone to large voids and shadow, so we propose a improved background subtraction algorithm based on segmentation. To divide the moving human images into two parts, choose appropriate threshold to binarization, then spliced into a complete image. Compared with the traditional background subtraction algorithm, the binary that extract is more accurate.(2) Gait Feature Extraction. Contrary to the traditional algorithm to extract gait feature is robustness to perspective and shelter innovation. This article purpose gait feature extract algorithm that based on focus track of the moving human and gait feature extract algorithm based on Poisson equation. The gait feature extract algorithm that based on focus track of the moving human calculate focus in a gait cycle to get focus track, by conduct FFT transform for focus track to eliminate the impact of clothing and belongings. FFT coefficients extracted as the focus track features. The gait feature extract algorithm based on Poisson equation propose a gait feature image G_F which can extract the protruding parts that difficult obscured(such as the head, legs, feet, etc) in any viewing angle. Using Gabor wavelet to characterization the gait feature image, using 2D-PCA to dimension reduction, get G_F feature. Experimental results show that both algorithm can be effectively eliminated the impact of clothing and belongings.(3) Gait recognition. To further enhance robust of algorithm to clothing and belongings on multiple views, improve the recognition rate, this article using feature-level fusion algorithm and match-level weighted addition fusion to fuse the focus track features and the G_F features of moving human. Focusing on feature-level fusion algorithm innovation, for traditional feature-level fusion algorithm that characteristic dimension of a substantial increase, proposing a feature parallel integration strategy, the formation of a parallel structure between the focus track features and the G_F features by introduction of complex vector, to research the fusion features in complex space. Compared to the two vectors and direct connect together to form a new vector, the new dimension feature vector is reduced, feature extraction and recognition speed. Eventually, select the Support Vector Machine(SVM) to make decision, validated by the CASIA Gait Database B of Chinese Academy of Sciences and own database, achieved good results, demonstrate the effectiveness of fusion methods.