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）》 2011年第5期515-523,共9页

摘　　要： 提出一种基于Laguerre正交基前向神经网络的动态手势识别方法.首先根据多项式逼近和矩阵理论,构造了一种以Laguerre正交多项式作为隐含层神经元激励函数的多输入、多输出三层前向神经网络模型,在网络权值迭代计算公式基础上推出一种基于伪逆的直接计算网络权值方法,避免求取权值的反复迭代过程;提出一种快速的基于颜色的指尖检测跟踪算法以便实时获取指尖运动轨迹,并提取指尖运动轨迹的特征向量作为Laguerre神经网络的输入向量;通过预先获取的动态手势样本(包括手势输入向量和预期结果)训练Laguerre神经网络,利用经过训练的Laguerre神经网络来识别通过摄像头获取的动态手势.测试结果表明:Laguerre正交基前向神经网络能够提高学习训练速度和精度,而且在动态手势识别方面具有较好的鲁棒性和泛化能力,具有较高的识别准确率. As a powerful tool for learning highly nonlinear system,BP neural networks are widely used in the field of nonlinear system identification such as pattern recognition.However,due to its shortcomings such as slow convergence rate,the danger of over fitting and low recognition accuracy,the traditional BP neural networks perform poorly in dynamic gesture recognition,especially for online dynamic gesture training.In this paper,a novel method of dynamic gesture learning and recognition based on Laguerre orthogonal basis feed-forward neural network was put forward.First,based on polynomial interpolation and matrix theory,a three-layer MIMO feed-forward neural network which hidden layer neurons are activated by a group of Laguerre orthogonal polynomial functions was constructed.In our research,the weights-updating formula for the neural networks was derived by adopting the standard BP training method,and then a pseudo-inverse based method which could determine the network weights directly without lengthy iterative training was proposed.Second,a rapid algorithm to detect and track the fingertips in real time based on machine vision was proposed.At first,a 2D color histogram on HSV color space was computed as the result of on-line training of fingertip colors,and it was used as the basis of fingertip detection and tracking;Then color probability distribution I（x,y） of the target was calculated through back-projection on the 2D color histogram and transformed into a binary image B（x,y） according to a given threshold;Then the target region of fingertips was determinated through edge detection and the centroid positions of targets could be calculated;At last,the trajectories of fingertips could be acquired by tracking the centroids of the fingertips through MDF（Minimum Distance First） and characteristic vectors of the dynamic gestures could be obtained.Third,some pre-obtained dynamic gestures that include characteristic vectors of gestrues and expected outputs were used as sample gestures to train the Laguerre ne

关 键 词： 正交基神经网络 权值直接确定 指尖跟踪 动态手势识别

领　　域： [自动化与计算机技术] [自动化与计算机技术]