A Real-Time Person-Following Framework Based on LiDAR-Camera Fusion.

Person-following technology is an essential part of the robot's human-machine interaction function and has a wide range of application scenarios. However, tracking a single person in the real world through sensors still faces many challenges, e.g., in dense crowds, the tracking target can be partially occluded or temporarily missed. In this paper, we proposed a robust person-following framework based on LiDAR-camera fusion to solve these problems. The framework consists of three main parts. Firstly, the person detection results are obtained from both sensors by using different detection methods. Secondly, a cascaded spatial-temporal data association method is used to filter out the currently tracked person. Finally, the target observations from both sensors are fed into different Kalman filters to update the target motion state in different observation coordinate systems. Experiments on KITTI and our tracking dataset show that, even under partial occlusion, our method can still track the target continuously. Moreover, the experimental results demonstrate that our method can be implemented in real-time with an average processing frame rate of 15 Hz.