基于深度图像的人手关节点识别-电子与通信工程专业论文.docx

摘 摘 要 江苏科技大学工学硕士学位论文 江苏科技大学工学硕士学位论文 万方数据 万方数据 万方数据 万方数据 摘 要 随着微软推出 Kinect 深度相机，基于深度图像的模式识别成为研究的热点，其中 符合人类交互习惯的手势识别成为研究的重要方向，人手关节点识别作为手势识别的 关键和基础问题，具有重要研究意义。深度图像每个像素点所带的信息，不是可见光 图像所带的亮度、色彩等信息，而是该点到相机的距离信息。与可见光图像相比，深 度图像不会受到光照、阴影、环境变化等因素的影响，并且直接反应了场景的三维信 息，为三维关节点识别提供了有利条件。 本文以基于深度图像的人手关节点识别为目标，提出了一套完整的解决方案。采 用深度相机捕获人手深度图像并通过真值标注方法制作深度图像样本库，提取样本的 深度特征用于随机森林模型的训练和部位识别，最后在各部位聚合像素信息得到关节 点的三维位置。本文设计的改进的深度比较特征，能有效区分人手的部位并且具有旋 转鲁棒性；采用的随机森林模型在识别问题上具有高效以及可以并行处理的优点。本 文的系统在PC平台上实现，对测试样本的部位识别的正确率为79%，识别的关节点与 标注关节点的位置偏差在允许范围内的概率达到82%，具有较高的识别准确率。本文 的算法平均处理一帧图像的时间是86毫秒，基本满足实时性要求。 关键词 深度图像；人手关节点识别；随机森林；改进的深度比较特征；部位识别 I II Ab Abstract 江苏科技大学工学硕士学位论文 江苏科技大学工学硕士学位论文 Abstract With Microsoft launching the K inect depth camera, depth image pattern recognition has become a research hotspot. Inside, hand gesture recognition has become an important research direction. As a key and basic problem, hand joint recognition of gesture recognition is of great research significance. The information of pixel in depth image is not the information of brightness or color in optical image, but the distance to the camera. Therefore, compared with the optical image, depth image will not be affected by the factors such as light, shadow, environment changes. Besides, depth image directly reflects the 3D information of the scene, which can provide favorable conditions for the three-dimensional joint recognition. This thesis puts forward a complete solution for hand joint recognition based on depth image. Capture hand depth image by the depth camera and use the method of labeling the real pixel to make the depth image sample database. Extract the depth features of samples for training random forest model and part recognition. Finally, combine information of pixels in each part to estimate the 3D position of joint point. The modified depth compared features designed in this article can effectively distinguish different hand parts and is of rotation robustness. Random forest model has the advantages of high effi