Graph-optimized coupled discriminant projections for cross-view gait recognition

Graph-Embedding is a widely used learning technique in pattern recognition. However, it is difficult to construct an inter-view graph for cross-view gait samples. To remedy it, in this paper, we propose a novel cross-view gait recognition algorithm named Graph-optimized Coupled Discriminant Projections (GoCDP), which seeks coupled projections based on adaptive graph learning. Regarding the embedding graphs as variables rather than predefined constants, we integrate inter-view graph construction with projection optimization process into a unified framework. By an alternate iteration algorithm, we can ultimately obtain the optimal coupled projections. Moreover, we extend GoCDP to multi-view case called Graph-optimized Multiview Discriminant Projections (GoMDP) for multi-view subspace learning. Experimental results on two benchmark gait datasets, CASIA-B and OU-ISIR, demonstrate the effectiveness of the proposed methods.