Person re-identification using view-dependent score-level fusion of gait and color features

This paper describes a method for person reidentification across multiple non-overlapping cameras using both gait and color features. Because a single color feature is insufficient to distinguish persons with similar color clothes, a spatio-temporal histogram of oriented gradients is employed as a gradient-based shape and motion gait feature to discriminate such persons in conjunction with a background edge attenuation technique. However, since the gait feature is more sensitive to view differences than the color feature, a view-dependent score-level fusion framework adaptively controls the weights of the gait and color features. Experiments across seven non-overlapping cameras confirm the effectiveness of the proposed method.