Research On Underwater Object Detection Algorithms For Underwater Vehicles

The deep blue sea has always been a place that arouses curiosity and imagination.Mankind has been trying to explore and exploit this mysterious part of the world fordecades. The advent of underwater vehicles improves our ability to understand theundersea world. The Underwater Vehicles, represented by the Remotely OperatedVehicle (ROV) and the Autonomous Underwater Vehicle (AUV), are being widelyused in scientific and military missions for search and survey, ocean engineering, testand evaluations, and in which the identification of underwater objects remains a majorissue.A number of underwater vehicles have used vision system as main sensingmechanism because of their high resolution and low cost. In order to get the broadvision, the wide-angle cameras are widely used in the Underwater Vehicles.However, the current domestic wide-angle vision system development also is notvery mature, for example, there is a problem such as image distortion usingwide-angle camera. In seawater, the light’s specific features such as absorption andscattering phenomena, make the vision task more difficult. In the specific context ofunderwater man-made object identification by robots, we often have some very basicknowledge on the object that we are looking for. Shapes and colors are the mostcommon knowledge. Because underwater imaging suffers from low contrast,non-uniform illumination, the literature related to underwater object identificationusually shows algorithms based on shape. Color is less frequently used as a feature forunderwater operations compared to terrestrial ones, but color remains a simple, robustand reliable piece of information for underwater object identification because of itssimplicity and its robustness to scale changes, object position and partial occlusions.Unfortunately, in underwater medium, the color is modified by attenuation and is not constant with the distance. To deal with color problems, some authors have also triedto embed suitable illumination systems or estimate with accuracy object surfacereflectance spectra. Those solutions work well but this kind of additional material oradditional prior knowledge is problematic constraint on light autonomous robots. Sohow to solve the wide-angle distortion and how to perform a color based detection ofan object are of great importance.This thesis is based on the vision system of ROV. In this thesis, on the basis ofcurrent research, the key problems of underwater object detection are focused, such aswide-angle distortion correction, the estimation of the attenuation parameters andcolor based detection of object algorithms. An improved x-corners detectionalgorithm on distorted images is proposed to improve the accuracy of the cameracalibration, and then to improve the correction of the wide-angle distortion. Colorbased detection of an object algorithm is implemented and improved. At last, anintegrated platform is introduced and tested. Several effective practices are doneunder water. The main achievements of research work are as follows:1) Aiming at the wide-angle distortion of image, this thesis presents animproved x-corners detection algorithm on distorted images. The algorithm is focusedon the accuracy of detecting of X corners on the calibration target, and this algorithmis based on gradients as the weights of center of gravity (COG). The results ofcorrection of the wide-angle distortion experiments show that the improved algorithmis more accurate and robust.2) The common color-based underwater object detection algorithms which mustadd material or add prior knowledge, is problematic constraint on light autonomousrobots. On the basis of the traditional classical model of color modification in anunderwater environment, the thesis then study the color-based underwater objectdetection algorithm using water light attenuation. Based on the current research, theattenuation parameters are estimated by using the sub images extraction from theimages corresponding to the object at different distances. And then aim to make thealgorithms more robust and to decrease the false alarms rate, we added some constraints. The underwater object detection experiments show that this algorithm isfast and more robust.3) This algorithm is implemented based on the Matlab toolbox. To test andverify the performance of the improved algorithm: firstly, the standard imagesprovided by Bazeille are used to do the object detection, and then the underwaterimages extracted using the wide-angle camera of ROV are tested to verify theperformance in the practical application environment. The experiments show that theimproved algorithm in this thesis improves the accurate and robust performance, andcan meet the demands for the Underwater Vehicles.