Using physics-informed machine learning to improve computer vision collision detection system.

Abstract

An enormous amount of the data humanity collects is visual data from monocular cameras. Despite this amount of data, there is a lack of research in the field on how the motion represented in these videos can be modelled in a physically accurate way. This thesis aims to develop a way to model optical flow data from an ordinary monocular camera with physics-informed machine learning and then to prove its effectiveness on the UAV collision detection problem. For this task, a Hamiltonian neural network is used to model optical flow measurements to predict a future trajectory of an object. These predictions are then used in a collision detection system that can work on data without any annotations. The approach is proven effective in modelling optical flow with a rigid physics-informed machine learning model. At the same time, a complex representation of the motion from several observations is proven to predict collisions accurately without sacrificing time to respond. The key takeaway from this study is that low-fidelity visual data, despite being an approximation of real-world motion, can describe fundamental physical properties in it.