针对正面步态识别中遇到背景颜色与人的穿着相接近或者人体下肢有阴影等情况，提出一种改进的背景减差法，利用髋关节将人体分为上下两个部分，采用分部阈值进行背景减差，提取人体二值化轮廓。对于周期检测，首先利用髋关节思想找到P点，然后利用P点找到人体左右脚最低点，并根据两脚间高低距离差所产生的角度计算步态周期并将其用于描述步态序列的动态特征；获得关键帧，并且进行归一化，分别提取它们的统一 Hu 矩和步态周期，用于描述步态序列的静态特征；将这两种特征进行融合，弥补了单一特征带来的识别率低的问题。最后，采用支持向量机(SVM)进行分类识别。本文的识别实验均在中科院步态数据库(CASIA)上进行训练，识别率达到97%以上。结果证明，该算法具有较好的识别速度和识别率。%Referring to both situations when the color of clothes people dressed is similar to the background color and when lower limbs of human body have the shadow in front-view gait recognition are hard to recognize, an improved background subtraction by using the different threshold at above and lower part of human body divided by the hip is proposed, which is used for extracting the binary contour of the human body. For cycle detection, first of all, we use hip ideas to find the optimal point P, and then obtain the lowest points of the left foot and right foot. The cycle is calculated by the angle based on the distance difference of the two feet, while the angle as a dynamic feature is used to describe the gait. The next step is to extract the key frames and be normalized, and the unified Hu moment features are extracted, being integrated with gait cycle and body proportion as the static features. The method of the fusion of static and dynamic feature solves the low recognition rate which is caused by the single feature. Finally, the Support Vector Machine (SVM) is used for classification. The recognition experiments of this paper are all trained in the Chinese Academy of Sciences gait database (CASIA), and the recognition rate is greater than 97%. The results show that the proposed approach has an encouraging recognition performance.
展开▼
