Underwater Change Detection by Fusing Multiple Sonar Images

Abstract: Underwater change detection can be used for monitoring the seafloor and automatically alert when something changes on it. This could be especially useful in harbors or other critical sea-infrastructure. The idea is to constantly survey the seabed with Autonmous Underwater Vehicles (AUV) equipped with different sonars such as sidescan and multibeam sonars. The data set used, consists of two subsets were one subset is recorded before any man-made objects had been placed on the seabed, the second subsets consists of sidescan data taken after the placement. The goal of this thesis is to develop a general approach to automatically detect changes on the seabed in general, and specifically try to find the man-made objects that were placed on the seabed. In order to achieve this, the thesis will mainly analyze sidescan data. The approach considered in this thesis is based on detecting objects in the sidescan data with a template matching algorithm. Then calculating the position of the objects in a global coordinate system and store all the objects positions from the first subset in a database. After that detect the objects in the second subset and compare their positions to the database. If the position of an object in the second subset does not exists in the database, that object can be considered a new object, and thus a change detection has occurred. Different template matching methods were tested and compared to each other with two test cases. Furthermore, preprocessing of the data were tested and compared as well. In order to calculate the objects position an optimized transformation between the global coordinates and the sonar’s frame of reference were calculated, the transformation is an important part of solving the problem since it will be necessary to do it for any kind approach to change detection. The template matching proved to be difficult to work in all scenarios, where it could only successfully detect all objects in the easier test case. Furthermore, the change detection proved not to be working due to the low success rate of the template matching. However, the change detection will work if the object detection is good enough.