Human Recognition by Appearance and Gait

Abstract

This work is focused on person identification task in video sequences. For this task we propose two complementing solutions, which can be applied in different cases: gait and visual recognition. For gait recognition three kinds of features are used: anthropometric features, based on the length of the skeleton segments; relative distance features, based on relative distances between the skeleton joints; and motion features, based on the movement of a joint between two frames. Two versions of the gait recognition algorithm are presented: the first one uses the depth data alongside with the images while the other one uses only the video sequence. For visual recognition from appearance we propose a deep learning algorithm that returns binary image features. Each algorithm was tested on two datasets. Furthermore, we perform experiments on transfer from one dataset to another to check trained model transferability.