Deep Person Re-identification with the Combination of Physical Biometric Information and Appearance Features

AbstractIn this paper, we propose a novel Person Re-identification model that combines physical biometric information and traditional appearance features. After manually obtaining a target human ROI from human detection results, the skeleton points of target person will be automatically extracted by OpenPose algorithm. Combining the skeleton points with the biometric information (height, shoulder width.) calculated by the vision-based geometric estimation, the further physical biometric information (stride length, swinging arm.) of target person could be estimated. In order to improve the person re-identification performance, an improved triplet loss function has been applied in the framework of [1] where both the human appearance feature and the calculated human biometric information are utilized by a full connection layer (FCL). Through the experiments carried out on public datasets and the real school surveillance video, the effectiveness and efficiency of proposed algorithm have been confirmed.