Occluded object tracking with RNNs
Mathis Petrovich · HAL (Le Centre pour la Communication Scientifique Directe) · 2018
Object detection in images has received a lot of attention in the recent years with state-of-the-art having human level performance. A lot of credit for such performance goes to the emergence of Deep Convolutional Neural Networks in particular and Neural Networks in general. Extending the capabilities to videos is a fundamentally more challenging task. Recent approaches in the field have mostly used a frame-based detector and linked those detections across time. A major problem with all existing work is the incapability to handle occlusion, as in cases of occlusion the detector fails to output a result and hence so does the linker. Such scenarios of occlusion are very common in human-robot-interaction where the robot has to track objects which might get occluded by the human hand. In this internship, I worked with a team of researchers with the aim of creating a tracking system which is robust to partial as well as total occlusion.