AI Powered Obstacle Distance Estimation for Onboard Autonomous Train Operation

Abstract

This paper proposes a novel method for an AI powered improvement of the estimation of a distance between the camera and an imaged object using image-plane homography. The method exploits the homography between two planes, the image plane and the rail tracks plane, and an artificial neural network that reduces the estimation error based on collected experimental data. The SMART multi-sensory onboard obstacle detection system has 3 vision sensors – an RGB camera, a thermal vision camera and a night vision camera, in order to achieve greater reliability and robustness. Although the methodology presented in this paper is applicable for each vision sensor, the proposed method was tested with the thermal camera and in impaired visibility scenarios. The validation of estimated distances is done with respect to real measured distances from the camera stand to the objects (humans) involved in the experiments. Distances are estimated with a maximum error of 2% and the proposed AI powered system can provide a reliable distance estimation in impaired visibility conditions.

Keywords

• Artificial neural network
• Autonomous train operation
• Distance estimation
• Homography
• Image processing
• Machine vision