Development of a machine-learning based object recognition system for quadrotors in urban environments

Abstract

This project presents the implementation of suitable Machine Learning (ML) architecture(s) to achieve real-time object detection and classification in a quadrotor in an urban environment, with a reasonable level of accuracy. Here, a suitable architecture refers to one that is able to achieve real-time performance, generally agreed to be 30 fps or higher among the community of ML practitioners. There is a compromise to be reached between accuracy and speed. Here, the constraint for speed is limited to the requirement of real-time performance. It is satisfactory to achieve levels of prediction accuracy comparable to current standards of reasonable accuracy, although the achievement of higher accuracy would be welcomed.