Detection and tracking of floating objects using a UAV with thermal camera

Abstract

This paper develops a vision-based tracking system in unmanned aerial vehicles based on thermal images. The tracking system are tailored toward objects at sea and consists of three main modules that are independent. The first module is an object detection algorithm that uses image analysis techniques to detect marine vessels in thermal images and extract the center of each object. Moreover, as long as the size of the vessel is known or computed in an image where the whole vessel is visible, the center can be identified in situations where only a part of the object is visible. The pixel position of the center is used in a nonlinear state estimator to estimate the position and velocity in a world-fixed coordinate frame. This is called the filtering part of the tracking system. The state estimator is nonlinear because only two coordinates in the world-frame can be computed with the pixel coordinates. This originates from the fact that cameras are bearing-only sensors that are unable to measure range. The last module in the tracking system is data association, which is used to relate new measurements with existing tracks. The tracking system is evaluated in two different case studies. The first case study investigates three different measures for data association in a Monte Carlo simulation. The second case study concerns tracking of a single object in a field experiment, where the object detection algorithm and the filtering part of the tracking system are evaluated. The results show that the modules in the tracking system are reliable with high precision.