基于运动与结构特征嵌入的人体行为识别

Human Behavior Recognition Based on Motion and Structure Feature Embedding

人体行为识别随着智慧城市等应用的普及逐渐受到重视。视频中的人体行为的表现主要体现在随着运动出现的像素的动态变化,以及在时空平面和空间平面中呈现出的结构信息上。受此启发,提出了基于运动与结构特征嵌入的人体行为识别方法。首先采用运动历史图像在二维平面上呈现三维运动变化信息,并结合三维运动的X-T、Y-T、X-Y平面的投影,构成7个平面来体现人体行为的结构信息。其次,利用高斯金字塔和中心-环绕机制,充分模拟人眼对尺度变化以及边缘敏感的特性,并结合Gabor滤波器的方向敏感性和最大池化的激励鲁棒性,提取人体行为特征表示。最后,采用判别局部对齐分析降维。在IXMAS数据集和真实的UCF Sports数据集上的实验验证所提出方法的有效性。

With the popularization of smart city and other applications,human behavior recognition has been paid more and more attention.The performance of human behavior in video is mainly reflected in the dynamic changes of pixels with motion,as well as the structural information presented in spatio-temporal plane and spatial plane.Inspired by this,a human behavior recognition method based on motion and structure feature embedding is proposed.The motion history image is used to present the three-dimensional motion change information on the two-dimensional plane,combined with the structural information reflected by the projection of X-T,Y-T and X-Y planes.Furthermore,Gaussian pyramid and center-surround mechanism are used to fully simulate the characteristics of human eyes,which is sensitive to scale change and edge.Additionally,combined with Gabor filter and max pooling,the approach is of directional sensitivity and excitation robustness.Finally,discriminant local alignment analysis is used to reduce the dimension,and excellent performances are obtained on IXMAS data set and UCF sports data set which can reflect real condition.