摘要： 在现有3D时空立方体的单个人体行为识别算法中,多数存在时间长度划分不准确和特征提取后降维处理时间复杂度高的问题。为此,提出一种基于人体运动关键点轨迹构成的3D立方体稀疏编码识别算法。定义人体的原子行为,将整个人体看做一个质点,分析其质点轨迹,并且根据人体原子行为的特性对轨迹进行分割,假设每段轨迹代表一个原子行为。将轨迹分段长度作为分割长度得到3D时空立方体,并由原子行为时间长度进行划分。取立方体每一帧中的关键点进行匹配,建立关键点轨迹。根据轨迹特性,通过排序对轨迹矩阵进行降维,避免复杂的矩阵运算,以降低时间复杂度。将降维后的轨迹矩阵进行稀疏编码,可得到不同运动方式下的稀疏表达。实验结果表明,与Blank、Castrodad等算法相比,该算法具有较高的识别率,且时间复杂度较低。

关键词: 行为识别, 3D时空立方体, 原子行为, 稀疏编码, 关键点轨迹

Abstract: In the existing human body behavior recognition algorithm for 3D space-time cubes,there is a problem that the time length is not accurately divided and the complexity of the dimension reduction processing time is high after feature extraction.Therefore,a 3D cube sparse coding recognition algorithm based on the trajectory of key points of human motion is proposed.Define the human body’s atomic behavior,the whole human body as a particle to get the particle trajectory,and according to the characteristics of the body’s atomic behavior on the trajectory segmentation,each trajectory represents an atomic behavior.The 3D space-time cube obtained by dividing the length of the trajectory segment as the segmentation length is divided by the atomic behavior time length.The key points in each frame of the cube are matched,and the key point trajectory is established.According to the characteristics of the trajectory,the dimension of the trajectory matrix is reduced by sorting,which effectively avoids complex matrix operations and reduces time complexity.Sparse coding of the dimensionality-reduced trajectory matrix is done,sparse expression under different motion modes is obtained.Experimental results show that compared with Blank,Castrodad and other algorithms,the algorithm has higher recognition rate and lower time complexity.

Key words: behavior recognition, 3D space-time cube, atom behavior, sparse coding, key points trajectory

中图分类号: 

• TP391