复杂环境下通用的手势识别方法

General method for gesture recognition in complex environment

用来解决手势识别中光照变化、背景干扰等问题的方法,往往计算量大,耗时长。针对这一问题,提出了一种复杂环境下通用的手势识别方法。该方法利用二进制方式实现支持向量机(SVM)模型并且使用位运算代替滑动窗口从而完成目标快速筛选,然后用统一计算设备架构(CUDA)实现卷积神经网络对初筛区域进行二次判断和识别。该方法不依赖于动态手势识别技术,可以应用于动态和静态的手势识别,能够同时处理光照变化、背景干扰的问题。实验结果表明所提算法的计算效率相比基于滑动窗口的算法有100至1000倍的提升,处理一幅图片的时间约为0.01 s。在修正后的Marcel数据集上实验结果达到了96.1%的准确率和100%的召回率。效率上的提升使得算法能够实时进行复杂环境下的手势识别。

The methods for dealing with influence of light and complex background often consume large calculation and long time. To solve this problem, a general method of gesture recognition in complex environment was proposed. The proposed method was based on the binary Support Vector Machine（ SVM） and bitwise operation instead of sliding window to achieve the goal of rapid screening, and then Compute Unified Device Architecture（ CUDA） was used to build a convolutional neural network to re-judge the initial screen area. The proposed method does not rely on dynamic gesture recognition techniques, and can be used for both dynamic and static gesture recognition. The method can deal with the problem of illumination change and background interference. The experimental results show that compared with the methods based on sliding window, the computational efficiency is improved by 100 to 1 000 times. It takes less than 0. 01 s to process a picture. The experimental results on the modified Marcel data set show that its precision achieves 96. 1% and recall achieves 100%. The proposed algorithm can be used for real-time hand gesture recognition under complex environment for its high performance.