电子测量与仪器学报

周宏宇,严春峰,宋旭,刘国英.基于加权三视角运动历史图像与时序分割的动作识别算法[J].电子测量与仪器学报,2020,34(11):194-203

基于加权三视角运动历史图像与时序分割的动作识别算法

Motion recognition based on weighted three-view motion history image coupled time segmentation

DOI：

英文关键词:motion recognition motion history image depth image sequence segmentation foreground contour self-occlusion

基金项目:国家自然科学基金(U1804153)、河南省高等学校重点科研项目(18B420001)、河南省高等学校重点科研项目计划(16B520001)、河南省科技计划(132102210212)、广西高校中青年教师基础能力提升项目(2019KY0223)资助

作者	单位
周宏宇	1. 安阳师范学院计算机与信息工程学院
严春峰	2. 桂林电子科技大学海洋工程学院
宋旭	1. 安阳师范学院计算机与信息工程学院
刘国英	1. 安阳师范学院计算机与信息工程学院

Author	Institution
Zhou Hongyu	1. School of Computer and Information Engineering, Anyang Normal University
Yan Chunfeng	2. College of Ocean Engineering, Guilin University of Electronic Science and Technology
Song Xu	1. School of Computer and Information Engineering, Anyang Normal University
Liu Guoying	1. School of Computer and Information Engineering, Anyang Normal University

摘要点击次数: 475

全文下载次数: 764

中文摘要:

针对当前人体动作识别算法中由于人体躯干遮挡而导致其检测精度不佳的问题,提出了一种基于加权三视角运动历史图像耦合时序分割的动作识别算法。首先,为了有效描述动作的形状和空间分布,从视频序列中提取运动历史图像(motion history image,MHI)。随后,应用深度相机(Kinect 相机)来提取深度图像,以获取人体目标的动作前景轮廓。为了识别由于身体部位造成的自我遮挡,动作前景轮廓被投影到 3 个视角(3V)平面,形成 3V-MHI,增强了对动作的正确提取,利用 3V-MHI 构造了一个用于记录观测运动轨迹的 MHI,克服了单视角 MHI 的信息局限性。然后,利用时序分割( temporal segmentation,TS),根据相邻的 3V-MHI 来计算动作的能量和方向的变化,以检测运动的开始和结束,从而输出运动结果。此外,计算 MHI 的梯度值作为每个平面对应的权重,从而得到加权 3V-MHI。最后,将提取的每个直方图运动模板与预先建立的数据库进行比较,完成动作的分类识别。实验表明,该方法能有效地解决自遮挡问题,在复杂环境和光照变化下有较高的准确性与鲁棒性。

英文摘要:

Aiming at the problem of poor recognition accuracy caused by human trunk occlusion in current human motion recognition algorithms, an action recognition algorithm based on weighted three-view motion history image coupling time series segmentation was proposed. Firstly, in order to effectively describe the shape and spatial distribution of the action, motion history image ( MHI) is extracted from the video sequence. Subsequently, the Kinect camera was used to extract the depth image to obtain the outline of the human target's action foreground. In order to recognize the self-occlusion caused by body parts, the outline of action foreground was projected to three view angles (3V) planes to form 3V-MHI, which enhances the correct extraction of action. Using 3V-MHI, a MII for recording and observing trajectories was constructed, which overcomes the information limitation of single-view MHI. Then, according to the adjacent 3V-MHI, the energy and direction of motion are calculated by using temporal segmentation to detect the beginning and end of motion and output the result of motion. In addition, the gradient value of MHI was calculated as the weight corresponding to each plane, and the weighted 3V-MHI was obtained. Finally, the extracted histogram motion template was compared with the pre-established database to complete the action classification and recognition. Experiments show that the method can effectively solve the problem of selfocclusion and has high accuracy and robustness for motion recognition under complex background and illumination changes.

查看全文查看/发表评论下载PDF阅读器