Monocular person tracking and identification with on-line deep feature selection for person following robots

This paper presents a new person tracking and identification framework based on solely a monocular camera. In this framework, we first track persons in the robot coordinate space using Unscented Kalman filter with the ground plane information and human height estimation. Then, we identify the target person to be followed with the combination of Convolutional Channel Features (CCF) and online boosting. It allows us to take advantage of deep neural network-based feature representation while adapting the person classifier to a specific target person depending on the circumstances. The entire system can be run on a recent embedded computation board with a GPU (NVIDIA Jetson TX2), and it can easily be reproduced and reused on a new mobile robot platform. Through evaluations, we validated that the proposed method outperforms existing person identification methods for mobile robots. We applied the proposed method to a real person following robot, and it has been shown that CCF-based person identification realizes robust person following in both indoor and outdoor environments.