The quality of the extracted gait silhouettes can hinder the performance and practicability of gait recognition algorithms. In this paper, we analyse the influence of silhouette quality caused by segmentation disparities, and propose a feature fusion strategy to improve recognition accuracy. Specifically, we first generate a dataset containing gait silhouette with various qualities generated by different segmentation algorithms, based on the CASIA Dataset B. We then project data into an embedded subspace, and fuse gallery features of different quality levels. To this end, we propose a fusion strategy based on Least Square QR-decomposition method. We perform classification based on the Euclidean distance between fused gallery features and probe features. Evaluation results show that the proposed fusion strategy attains important improvements on recognition accuracy.
展开▼
