Adaptive local module weight for feature fusion in gait identification

Abstract

In partial occlusion problem, we have many methods for gait identification. However, based on Gait Energy Image (GEI) part, they are hardly occluded local part threshold problem. We propose an adaptive local module weight to reduce score of the occluded part. The adaptive local module weight is constructed from a consensus and a complementary principles of global and local modules. Firstly, we construct the consensus principle from a row reliability and an accuracy identification weights. Then, the complementary principle is constructed from a shape weight. We extract all features by the combined TDPCA and TDLDA method. The similarity of testing module and all modules in gallery are measured by the Euclidean distance. Finally, we combine the row reliability, accuracy identification, and shape weights with the similarity scores for gait identification. For evaluating our proposed method, we use the silhouette image sequences from the EEPIT dataset with 135 classes and the CASIA dataset with 123 classes. The results show the recognition effectiveness of the proposed method over than the conventional method.