Deep Learning for Long and Short Range Object Detection in Underwater Environment
Docking capabilities are required for long operation of an Autonomous Underwater Vehicle (AUV) due to limited battery and data storage capacity. Docking is composed of four main processes, i.e. homing, plug-in, release and drive-out. The homing part requires detection and localization of the docking station, guidance and control to reach the docking station. Hence, in this paper, we focus on the process of detection and localization. The AUV is assumed to use two sensors for perception of the environment, i.e. an imaging sonar and a monocular camera which is used in close range navigation. Deep learning (DeepL) methods are well-known for good detection and localization. Therefore, in this paper, two DeepL networks will be designed. The first one is for detecting the target object in a sonar image at far range and the other one is for detection of the docking station in the close range using data from an optical camera image. The results from experimental studies in a test basin with an AUV show that the proposed system is able to locate and classify the docking station in both optical and sonar images with detection rate of 94.3% and 80%, respectively.