Real Time Semantic Segmentation with SLAM for Gaze Intention Decoding


Understanding the human gaze’s role in conveying intention is an important step in developing intuitive controls for human-in-the-loop robotics which could aid those who are physically disabled, among many other applications which involve direct human interaction. Recent work has primarily focused on studying gaze using a 2D ego-centric view of a screen. However, this does not reflect reality, where the 3D environment around us affects our gaze. To decode 3D gaze intention from an ego-centric perspective, we need to know what the subject is looking at, how the subject is looking at it and where that object is in relation to the user in 3D space. In this work, we use recent improvements in the fields of object detection, semantic segmentation and simultaneous localisation and mapping to build a system which can both map and label the 3D environment in real-time through RGB-D data from an ego-centric point of view. This has resulted in a novel sparse semantic/instance segmentation method, based on the YOLO object detection network, which we have titled “YoloMask”. After initial training, YoloMask performs 7 times faster at inference than the state of the art. Following this we also present an extended version of SemanticFusion, an existing semantic SLAM algorithm, by incorporating our YoloMask network into it, expanding the number of detectable classes from the original 17 to 80, while providing a speed increase allowing it to run in real-time. In line with the gaze intention detection aim of this work, we also built a modular gaze tracking tool which supports commercial eye-tracking glasses and is used as an experiment runner to collect large amounts of data on human gaze behaviour. This report documents the development, implementation, training and evaluation of each of these systems. The outcomes of this work will hopefully provide the building blocks for further research into 3D gaze intention decoding.

Department of Computer Science, Imperial College, London

Joint Mathematics and Computing Integrated Masters Thesis