Part-Based Pedestrian Attribute Analysis

Visual pedestrian attributes are very important for person re-identification. Due to the difficulties in obtaining identifiable face and body shots in surveillance scenarios, clothing appearance attributes become the main cue for identification. In this paper, we propose a part-based pedestrian attribute analysis method upon clothing appearance. First, to alleviate pose misalignment, 25 anatomical key-points are located by OpenPose algorithm and then 9 body parts are extracted. Besides, 3 poses are recognized via constraints on key-points. Second, for each part image, the main color feature is extracted by ColorName algorithm, and the texture classification feature is extracted by CNN network combined with SVM model. Finally, for pedestrian pair with certain poses, a weighted similarity fusion algorithm based on the color and texture feature is applied to compute the total similarity of two sets of body parts. Experimental results on pedestrians in surveillance videos demonstrate the effectiveness of our method.