Research On Key Techniques For Gait Recognition

Gait recognition has recently attracted much attention in the areas of pattern recognition and computer vision. Gait recognition, aiming at an authentication based on walking style, is currently the only one biometric technology that can achieve a long distance identification. Compared with other biometric technologies, gait recognition has many advantages. For example, it can fulfill a contactless identification from a long distance and demands lower video quality. Moreover, gait is unique to individuals and difficult to camouflage. Therefore, it is important and promising to study gait recognition.This work addresses a few problems that have long been in the heart of gait recognition: how to increase the recognition accuracy and robustness and how to reduce the computation complexity to guarantee a real-time performance. The major work is as follows:1) We conduct a comprehensive comparison among the average model, the median model and the Gaussian mixed model. We propose an improved method for background update which is based on the standard deviation. The proposed method is more efficient in computation and leads to less background noise, enabling a real-time recognition of the gradual change in background. Rather than a fixed threshold, we adopt a double thresholds mechanism which is adaptive to the mean and the standard deviation. A more ideal binary moving targets can be obtained with our method.2) We propose a novel gait cycle generation method based on the fusion of multiple cycles. The proposed method pools those high quality frames from each original gait cycle to compose a final cycle output, producing a more accurate gait cycle especially when the original sequence contains incomplete and duplicate frames. This method can not only eliminate the influence of changing speed or pause of the moving target, but also reduce the data redundancy and the computation cost.3) We summarize the characteristics of gait energy image(GEI) and point out that the time information carrying individual characteristics is missed in the GEI. To reduce data storage, we propose a gait compressed image. The proposed gait compressed image keeps the frequency as well as the time information of a gait. Two-dimensional principal component analysis and linear discriminant analysis are adopted in composing the gait compressed image. Finally by observing the experimental process, the different part of the individual contributes to the recognition rate is not the same, and then begin the experiment and analysis. It may provide a new way of research on gait recognition in different situation.Based on our research, we have found:1) If the foreground color is monotonous and uniform, the adaptive double thresholds method based on the mean and standard deviation leads to a gait sequence with fewer incomplete frames and noise. 2) When a gait sequence contains duplicate or incomplete frames, gait cycle generation based on the fusion of multiple cycles can eliminate the influence of the unqualified frames and make full use of the ideal frames in all of the gait periods, so that a more accurate gait cycle can be achieved. 3) Head barycenter registration method results in a better gait energy image. 4) Gait compressed image is an effective and feasible method of gait sequence characterization, retaining both the frequency and the time information of the gait sequence. 5) When the sum of the eigenvalues’ square in the covariance matrix is greater, the recognition effect is better. And the recognition rate will no longer increase while the retained eigenvalues number increases to a certain extent. 6) Different parts of the individual contributes differently to the recognition result. The body can easily be disturbed and thus results in a lower recognition rate, while the head and foot contribute most to the recognition result.