DISTANCE ESTIMATION MODEL FOR THERMAL VISION SYSTEMS USING GAUSSIAN PROCESS REGRESSION

Abstract

An obstacle detection system (ODS) that can operate in a challenging environment and with limited visibility is the crucial element of autonomous systems. Distance estimation of the objects (obstacles, humans) plays a very important role in each ODS. In this paper, to estimate the distance between the camera and objects, a Gaussian processes regression (GPR) model was proposed. GPR is a machine learning method that is based on Bayesian and statistical learning theory. GPR is an effective method for processing data and predicting/estimating information that is adaptable to complex regression problems such as high dimensions, small samples, and non-linearity. The proposed model's data were collected for training and testing using the Smart Automation of Railway Transport (SMART) onboard obstacle detection system developed to achieve autonomous train operation (ATO). The bounding box features of the recognized object are used as input data, while output data was obtained by measuring distances form the camera and humans involved in the experiment. The GPR results are compared with measured distances and distance estimations obtained by image-plane homography. The proposed method was tested with the thermal camera and in impaired visibility scenario, but the presented methodology could be applied for various vision sensors.