基于动态时间规整耦合3D运动历史图像的人体动作识别算法  

Human Motion Recognition Algorithm Based on Dynamic Time Warping Coupled With 3D Motion History Image

作　　者：石峰[1] SHI Feng(Department of Computer Science and Technology,Taiyuan University,Taiyuan Shanxi 030032,China)

机构地区：[1]太原学院计算机科学与技术系,山西太原030032

出　　处：《传感技术学报》2024年第11期1937-1945,共9页Chinese Journal of Sensors and Actuators

基　　金：山西省高等学校改革创新项目基金项目(J20221192)。

摘　　要：针对当前动作识别过程中忽略了场景的语义信息,易受视角变换与遮挡的影响,导致识别率不高等问题,提出了一种基于动态时间规整耦合3D运动历史图像的人体动作识别算法。首先,结合人体的空间位置、运动方向和速度等不同特征,利用多维最长公共子序列(Multi-Dimensional Longest Common Subsequence,MDLCS),对视频数据中的行人目标进行跟踪,提取目标的运动轨迹。然后,基于频谱映射理论,对得到的轨迹实施聚类,并计算运动轨迹的聚类中心。通过对聚类结果执行ROI划分和提取,获取场景的语义上下文信息。再引入动态时间规整(Dynamic Time Warping,DTW),将输入的视频序列与聚类中心进行比较,消除异常与冗余动作信息。随后,计算轨迹段的起点、终点与工作区的ROI之间的位置关系,结合场景的语义上下文信息,采用基于颜色和深度信息的3D运动历史图像(3D Motion History Image,3D-MHI)来提取动作特征。最后,利用支持向量机(Support Vector Machine,SVM)对3D-MHI动作特征进行分类学习,完成对人体动作的识别。实验表明:所提算法在UCF Sport与Hollywood数据集上的识别率分别达到了95.1%和92.5%,与当前流行的动作识别算法比较,具有更高的识别率与较强的鲁棒性,对视角变换与遮挡等复杂场景下的动作识别更为有效。In view of the fact that the semantic information of the scene is ignored in the current motion recognition process,which is easily affected by the angle of view transformation and occlusion,resulting in the low recognition rate,a human motion recognition scheme based on dynamic time warping coupled with 3D motion history image is proposed.Firstly,the multi-dimensional longest common subsequence(MDLCS)is used to track the pedestrian target in the video data and extract the moving track of the target.Then,based on the spectrum mapping theory,the trajectory is clustered and the clustering center of the motion trajectory is calculated.Through ROI partition and extraction of clustering results,semantic context information of scene is extracted,which provides important constraints for subsequent action analysis.Then,dynamic time warping(DTW)is introduced to compare the input video sequence with the clustering center to eliminate the abnormal and redundant action information.In order to improve the semantic information of the scene,the position relationship between the starting point and the end point of the track segment and the ROI of the workspace is calculated.the 3D motion history image(3D-MHI)based on color and depth data is combined with the semantic context of the scene to extract action features.Finally,SVM is used to classify and learn 3D-MHI motion features to complete the recognition of human motion.The experimental results show that the recognition rates of the proposed algorithm are 95.1%and 92.5%respectively.Compared with the current popular action recognition algorithm,the proposed method has higher recognition rate and stronger robustness.It is more effective under the circumstance of view transformation and occulusion.

关 键 词：动作识别 动态时间规整 多维最长公共子序列 频谱映射 3D运动历史图像 聚类中心 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]