Multi-feature Fusion Of Human Gait Recognition In Complex Background

With the growing popularity of intelligent video surveillance technology, there is an urgent requirement for more accurate human gait recognition on movement of the target area.This paper studies the human gait recognition with multi-feature fusion,which under complex background conditions (for example, leaves disturbances, changes in illumination, etc.).Experiments prove that the Gaussian mixture background modeling adapt to complex background (such as leaves disturbances, changes in illumination, etc.) poorly,but the K-means clustering image segmentation has good robustness on light and leaves tremble, and has little chance of resulting empty divided image. The K-means clustering image segmentation algorithm which based on micro-canonical annealing was proposed through in-depth analysis of pixel-level characteristics of the Gaussian mixture background modeling, K-means clustering characteristics of the image region segmentation and efficient global optimization feature of micro-canonical Annealing, so that the interested region obtained by dividing become more accurate and complete. Then, an improved Gaussian mixture background modeling on image segmentation appears.Bidirectional 2DPCA compresses on rows and columns during feature extraction, so that the amount of data and computation will reduced, this is why it’s been used to extract characteristic from gait energy image. Traditional joint angle characteristic is too idealistic, which obtain lower limb joints through ratios between limb and height of human anatomy, and this paper proposes a new method to get joint angles based on skeleton model to remove the tip points, whose result will be more accurate. Discrete Hu moment invariants with translation, scaling and rotation invariance can be a good representative of static information on gait. Meanwhile, three kinds of gait characteristics are so different on properties and numbers that normalization is really needed.The fusion method on weighted-feature level without feedback is been used considering complexity on computing and complementarities of characteristics, which obtain a recognition rate on single feature firstly,then get weights of each feature,and adjust the weights at last. Support vector machine has a unique advantage in recognition of small sample, nonlinear and high-dimension pattern, so it is been used to feature classification. During experiments, on one hand, support vector machine will classify single or fusion characteristics to get respective recognition rate, which greatly improve the multi-feature fusion recognition algorithm, on the other hand, comparison analysis on cumulative matching score will make an evaluation to classification performance of the improved multi-feature fusion gait recognition algorithm.