为了改善常见的背景模型在手势分割易受环境因素影响的缺点, 以及提高基于傅里叶描述子的神经网络对大量手势分类的识别率, 本文提出一种采用预分类的综合手势识别算法。首先, 分割部分使用背景差分检测出手部, 同时利用手分类器对手部识别手掌位置, 获取感兴趣区域, 同时对手和背景分别进行掩膜更新, 进入下一循环, 具有较高的分割效果。接着, 利用指尖检测和傅里叶描述子对区域进行特征提取。识别过程先根据指尖数目及实现功能进行预分类, 再利用神经网络对傅里叶特征系数进行实时的识别并返回识别结果。最后, 进行与直接的傅里叶系数的识别比较及复杂背景下与传统背景差分方法识别率比较, 结果表明该算法普通手势比直接分类平均识别率高4.162%, 易误检手势识别率提高最高达91.7%; 复杂背景下性能也有较大改善。鲁棒性较强, 可以容纳大量手势定义, 满足人机交互的要求。

To improve the common background model that can be easily faulted by environmental factors, and the efficiency of neural network based on Fourier descriptor for a large number of gesture recognition, a pre-classified gesture recognition algorithm is designed. First, background difference detection plus a hand classifier is employed to recognize palm position and obtain regions of interest, while masks are updated respectively in hand regions and background. Experiments indicate its remarkable segmentation effects. Then, fingertips detection follows and Fourier descriptors are extracted from areas of interest. Identification process includes the preliminary classification dependent on the number of fingertips detected as well as the potential movement. Then a neural network based on Fourier coefficients is to process the extracted feature data and feedback the results. In the end, comparison shows that this algorithm involving pre-classification system performs better in recognition. Ordinary gestures improve 4.162% in average, while easy-mistaken gestures escalate by up to 91.7%. Under complex environment, improvement can be seen compared to a common background difference method. Strong robustness and multiple gestures definition satisfy the requirement of real-time human-computer interaction.