Vision Based Detection and Tracking in Dynamic Environments with Minimal Supervision

This thesis presents vision based object detection and tracking techniques suitable for dynamic and outdoor applications with a moving camera. Firstly, a motion clustering approach is presented to discover dynamic objects with previously unknown appearance and then used to train an appearance based model. Secondly, a novel background appearance model is proposed to verify the output of a pretrained deep convolutional network based object detector. The combined detector is demonstrated to significantly improve the pretrained detector with only weak supervision from background images when transferred to a mine site environment. Finally, a framework for associating detections across frames is presented that exploits spatial and temporal constraints, enabling life-long improvement through self-supervised learning.