Active Vision System with Human Detection - Using RGB-D images and machine learning algorithms

Abstract

This master's thesis will focus on an active safety system for the protection of humans close to commercial construction equipments. The purpose is therefore to propose sensors and algorithms suitable for human detection and furthermore to demonstrate a proof of concept. Early on in the project it was decided to use RGB-D images, which is a conventional color image together with a depth map. This report analyzes both a Kinect sensor and a stereo vision system in order to generate a depth map. Machine learning algorithms were used to classify humans where an artificial neural network was found to be the best performing classifier. Finding informative features is important to facilitate classification. Several imaging features were tested and the six most interesting are presented in this report. The feature called fourier descriptor showed the best performance.