作　　者： ; ; ;

机构地区： 中山学院计算机学院

出　　处： 《南京大学学报（自然科学版）》 2012年第4期421-428,共8页

摘　　要： 针对动态手势识别中传统神经网络训练算法存在收敛速度慢、网络精度低等缺陷,提出一种自适应MIMO-Chebyshev神经网络(MIMO-CNN)算法实现动态手势学习和识别.以Chebyshev正交多项式作为隐含层神经元激励函数构造多输入、多输出三层前馈神经网络,并给出权值直接确定方法和隐含层节点数目自适应确定算法.给出一个基于颜色直方图的指尖检测算法和基于二部图最优匹配的指尖跟踪算法以便实时获取动态手势轨迹.针对动态手势识别要求对MIMO-CNN进行输入输出结构设计和网络权值学习训练,并运用经过训练的MIMO-CNN识别动态手势.测试结果表明:MIMO-CNN能够提高网络训练速度和精度,从而提高动态手势学习速度和识别准确率,而且在动态手势识别方面具有较好的鲁棒性和泛化能力. Due to its shortcomings such as slow convergence rate, the danger of over fitting and low network accuracy, the traditional back propagation（BP） neural networks perform not very well in dynamic gesture learning and recognition. In this paper, a novel adaptive MIMO-Chebyshev neural networks（MIMO-CNN） algorithm for dynamic gesture learning and recognition is proposed. First, a feed-forward neural network which hidden layer neurons are activated by a group of order-increasing Chebyshev orthogonal polynomial functions is constructed. As Chebyshev function is an unary function and the Chebyshev neural network is convergent only when the input is in [-1,1], we adapt a S-function to transform the input from R to.Rl and from （-∞∞） to [-0.1]. Based onpolynomial interpolation, approximation and matrix theory, the weights-updating formula for such a neural network is derived by adopting the standard BP training method, and then a pseudo-inverse based method which can determine the network weights directly without lengthy iterative training is proposed. By adopting weight direct- determination method, we also present an auomatic-determination algorithm for the determination of hidden-layer neuron number of MIMO-CNN according to the accuracy goal. Second, a rapid algorithm to detect and track the fingertips in real time is discussed. A rapid method to detect the fingertips based on color histogram on HSV color space（H and S） is put forward at first. However, because the color histograms and the shape characteristics of fingertips are very similar and multiple fingertips may cross, traditional target tracking algorithms such as CamShift, Kalman filtering, etc. may fail in the fingertip tracking. Considering the idea that the tracking of fingertips can be transformed into a fingertip-match problem between consecutive frames, we propose a method to implement fingertip tracking based on Optimal Matching Algorithm of Bipartite Graph. Then, the traiectories of fingertips can be acquired by tracking fingertips an

关 键 词： 神经网络 手势跟踪 手势学习 手势识别

领　　域： [自动化与计算机技术] [自动化与计算机技术]