Real-Time Online Human Tracking with a Stereo Camera for Person-Following Robots

Person-Following Robots have been studied for multiple decades now. Recently, person-following robots have relied on various sensors (e.g., radar, infrared, laser, ultrasonic, etc). However, these technologies lack the use of the most reliable information from visible colors (visible light cameras) for high-level perception; therefore, many of them are not stable when the robot is placed under complex environments (e.g., crowded scenes, occlusion, target disappearance, etc.). In this thesis, we are presenting three different approaches to track a human target for person-following robots in challenging situations (e.g., partial and full occlusions, appearance changes, pose changes, illumination changes, or distractor wearing the similar clothes, etc.) with a stereo depth camera. The newest tracker (SiamMDH, a Siamese convolutional neural network based tracker with temporary appearance model) implemented in this work achieves 98.92% accuracy with location error threshold 50 pixels and 92.94% success rate with IoU threshold 0.5 on our extensive person-following dataset.