Data-driven Obstacle Detection from Dense Optical Flow

Lead Research Organisation: Heriot-Watt University
Department Name: Sch of Engineering and Physical Science


Obstacle avoidance with monocular vision is an important open problem in robotics. Measurements of optical flow (OF) carry information about potential collisions directly without the need to identify complex and abstract semantics of the video stream. However, dense OF for navigation remains a poorly explored area because, until recently, real-time dense OF estimation was simply not available. This project was aimed at harnessing the power of modern neural networks to extract collision information from dense OF of a robots video stream. A pixel-wise classifier was proposed to identify regions of high collision probability. This exploratory work discusses potential solutions to challenges specific to OF frames processing, as well as high class imbalance of the collected training data.


10 25 50

Studentship Projects

Project Reference Relationship Related To Start End Student Name
EP/R512539/1 01/10/2017 30/09/2021
1944026 Studentship EP/R512539/1 01/09/2017 31/08/2022 Jakub Sanak