ALBERT

Description: The scope of the project described in this paper is the development of a generalized underwater object detection solution based on Automated Machine Learning (AutoML) principles. Multiple scales, dual priorities, speed, limited data, and class imbalance make object detection a very challenging task. In underwater object detection, further complications come in to play due to acoustic image problems such as non-homogeneous resolution, non-uniform intensity, speckle noise, acoustic shadowing, acoustic reverberation, and multipath problems. Therefore, we focus on finding solutions to the problems along the underwater object detection pipeline. A pipeline for realizing a robust generic object detector will be described and demonstrated on a case study of detection of an underwater docking station in sonar images. The system shows an overall detection and classification performance average precision (AP) score of 0.98392 for a test set of 5000 underwater sonar frames.

Description: The utilization of stationary underwater cameras is a modern and well-adapted approach to provide a continuous and cost-effective long-term solution to monitor underwater habitats of particular interest. A common goal of such monitoring systems is to gain better insight into the dynamics and condition of populations of various marine organisms, such as migratory or commercially relevant fish taxa. This paper describes a complete processing pipeline to automatically determine the abundance, type and estimate the size of biological taxa from stereoscopic video data captured by the stereo camera of a stationary Underwater Fish Observatory (UFO). A calibration of the recording system was carried out in situ and, afterward, validated using the synchronously recorded sonar data. The video data were recorded continuously for nearly one year in the Kiel Fjord, an inlet of the Baltic Sea in northern Germany. It shows underwater organisms in their natural behavior, as passive low-light cameras were used instead of active lighting to dampen attraction effects and allow for the least invasive recording possible. The recorded raw data are pre-filtered by an adaptive background estimation to extract sequences with activity, which are then processed by a deep detection network, i.e., Yolov5. This provides the location and type of organisms detected in each video frame of both cameras, which are used to calculate stereo correspondences following a basic matching scheme. In a subsequent step, the size and distance of the depicted organisms are approximated using the corner coordinates of the matched bounding boxes. The Yolov5 model employed in this study was trained on a novel dataset comprising 73,144 images and 92,899 bounding box annotations for 10 categories of marine animals. The model achieved a mean detection accuracy of 92.4%, a mean average precision (mAP) of 94.8% and an F1 score of 93%.