基于联动空间拓扑关系的步态识别研究

Research on gait recognition based on spatial topology and multi⁃scale temporal features

目前,大多数基于骨骼的步态识别方法通过改善时空关系表征来提高步态识别的准确率,但却忽略了人体物理结构节点之间的联动性。基于上述问题,探究步态识别的空间拓扑联动关系。提出一种基于局部分区的新策略方法,划分人体非结构化区域并构建区域之间的邻接关系表示人体姿态的相对变化。其次,设计多尺度时间模块以时间为轴关联局部分区的节点变化提取多尺度时间特征。通过实验验证,该方法在步态数据集CASIA-B上超越了主流的步态识别方法,在不同行走条件下Rank-1准确率分别达到87.6%、77.6%以及72.8%。

Currently,most of the skeleton⁃based gait recognition methods improve the accuracy of gait recognition by improv⁃ing the representation of spatiotemporal relationship,but ignore the linkage between the nodes of human physical structure and the cues of long and short time series.Based on the above problems,the spatial topological linkage of gait recognition is explored.In this paper,a new strategy method based on local partitioning is proposed to divide unstructured regions of the human body and con⁃struct the adjacency relationship between the regions to represent the relative changes of human posture.Secondly,a multi⁃scale temporal module is designed to extract multi⁃scale temporal features by associating node changes of local partitions with time as the axis.Through experimental verification,the proposed method surpasses the mainstream gait recognition methods on gait dataset CASIA⁃B,and the Rank⁃1 accuracy rate reaches 87.6%,77.6%and 72.8%respectively under different walking conditions.