The Gait Feature Extraction And Recognition Based On Machine Vision

Walking is one of the basic human motions, so the gait recognition technique can identify people by their walking pattern. Gait is the only distance captured biological characteristic, it has gained increasing interest in biometric identification field, especially for intelligent visual surveillance at a distance and monitoring systems in security-sensitive environment. The study of gait recognition has promoted the development of the theory of digital image processing, machine vision, pattern recognition and signal processing, so it was important for scientific research. Generally, gait recognition consists of three parts:gait detection, feature extraction and recognition. Concerning the three parts, this paper discussed several themes as follows:Firstly, according to motion detection methods and our actual situation, we chose the background subtraction to get the human body upper and lower limbs articulation point in HSL color model. Threshold segmentation, morphologic and particle removal operations were used to make the image binarization; these jobs laid a good foundation for the following feature extraction works.Secondly, on the feature extraction aspect, we used lower-limb joint angle to divide period, chose the motion characteristics of upper and lower limbs. Fourier series fitting was used to extract features of lower-limb amplitude spectrum; together with the discrete Fourier transform analyzed the vertical components which are the wrist joint z-coordination, to get the amplitude spectrum features. Experimental results demonstrate that the feature of amplitude spectrum is more stable than the space characteristics.At last, according to the idea of information fusion, a gait recognition method using weighted average fusion of upper-limb wrist joint trajectory z coordination and lower-limb angles at feature level was proposed. Weighted average method was used to make each feature assigned to weights, which can make them combine in suitable proportion.Experimental results demonstrate that, to a certain extent, the method of this paper gets some improvements in all aspects, such as recognition rate is more than97%.