查询结果:   LE Vanbang,朱煜*,NGUYEN Anhtu.深度图像手势分割及HOG-SVM手势识别方法研究[J].计算机应用与软件,2016,33(12):122 - 126.
中文标题
深度图像手势分割及HOG-SVM手势识别方法研究
发表栏目
人工智能与识别
摘要点击数
784
英文标题
RESEARCH ON DEPTH IMAGE GESTURE SEGMENTATION AND HOG-SVM GESTURE RECOGNITION METHOD
作 者
LE Vanbang 朱煜* NGUYEN Anhtu LE Vanbang Zhu Yu* NGUYEN Anhtu
作者单位
华东理工大学信息科学与工程学院 上海 200237     
英文单位
School of Information Science and Engineering,East China University of Science and Technology,Shanghai 200237,China     
关键词
手势分割 深度图像 距离变换 HOG 支持向量机
Keywords
Gesture segmentation Depth image Distance transform HOG Support vector machine
基金项目
国家自然科学基金项目(61271349);中央高校基本科研业务费专项资金项目(WH1214015)
作者资料
LE Vanbang,博士生,主研领域:图像处理,模式识别,手势识别。朱煜,教授。NGUYEN Anhtu,硕士生。 。
文章摘要
针对深度图像静态手势识别问题,提出一种基于深度图像手势分割及HOG-SVM手势识别方法。该方法的具体做法包含以下四个步骤:第一步,对深度图像进行手势分割,对随机方向的手臂图像通过椭圆拟合算法计算其倾斜角度,并将其校正至垂直方向;第二步,对手臂图像进行距离变换,通过分析距离变换返回的距离矩阵精确定位手掌心、手腕及手臂在图像中的坐标;第三步,计算、优化手势图像的HOG特征;第四步,实时采集大量训练样本并获取其训练矩阵,对训练矩阵进行处理找到最优的SVM参数,使响应曲线的可区分度达到最佳以提高手势识别率。实验证明,所设计的系统在保证实时性、鲁棒性的同时也获得了很高的识别率。
Abstract
Aiming at the static gesture recognition of depth image,this paper proposes a depth image-based gesture segmentation and HOG-SVM gesture recognition method.The specific approach of the method contains following four steps.First,we carry out hand gesture segmentation on depth image and calculate tilt angle of the arm image in stochastic direction by ellipse fitting algorithm,and then regulate it to vertical direction.Secondly,we make distance transform on arm image,and precisely locate the coordinates of palm,wrist and arm in the image by analysing the distance matrix returned from distance transform.Thirdly,we calculate and optimise the HOG features of gesture image.Finally,we collect in real time a large number of training samples and obtain their training matrix,process the training matrix to find the optimal SVM parameters,thus enabling the distinguishable degree of the response curve to reach the best so as to improve gesture recognition rate.It is proved by the experiment that the designed system in the paper achieves quite high recognition rate while ensuring real-time performance and robustness.
下载PDF全文