Detecting Water Reflection Symmetries in Point Clouds for Camera Position Calibration in Unmanned Surface Vehicles

Li Qingqing, Jorge Peña Queralta, Tuan Nguyen Gia, Zhuo Zou, Hannu Tenhunen and Tomi Westerlund

The development of autonomous vehicles has seen considerable advances over the past decade. However, specific challenges remain in the area of autonomous waterborne navigation. Two key aspects in autonomous surface vehicles are sensor calibration and segmentation of water surface. Cameras and other sensors in a car or drone can be installed accurately in a specific position and orientation. In a large vessel, this is not always possible, as sensors might be installed around the vessel or on masts. Taking advantage of the medium in which these vehicles operate, the water plane can be used as a reference for different sensors to calibrate their orientation. This allows more accurate localization of obstacles of objects. State-of-the-art deep learning techniques have been successfully applied for water surface segmentation in open sea. However, in other environments such as small rivers or lakes with still waters, a different approach might enable more accurate water surface estimation. We propose a method to estimate the water plane based on the detection of local symmetry planes that naturally occur when objects are reflected in the water. By using a point cloud generated with a stereo camera, we are able to accurately estimate the water level and, at the same time, calibrate the camera position and improve the localization of obstacles. We assume that an approximate position and orientation of the camera is known with respect to the sea level. We demonstrate the efficiency of our method with data obtained in the Aura river, Finland, with our prototype vessel facing both the riverside and the center of the river.