Activity Segmentation and Fish Tracking From Sonar Videos by Combining Artifacts Filtering and a Kalman Approach
Julian Winkler, Sabah Badri-Hoeher, Fatna Barkouch
Abstract
Fish stocks are among the most endangered components of marine ecosystems. To minimize threats to marine ecosystems and ensure natural and sustainable resource use, monitoring systems must be placed in oceans and seas. The Underwater Fish Observatory (UFO) and the UFOTriNet are two projects initiated by several researchers from marine biology, engineering and industry in Germany between the years 2014 and 2016 and between 2019 and 2023, respectively. The systems collect abiotic as well as camera and sonar data to count fish stocks. This work proposes a method for robust fish counting using sonar data. Activity segmentation as well as object tracking are important steps to successfully accomplish this task. Background subtraction is often used as pre-processing step for stationary fixed sonars. Our proposed method improves this step using band-pass filtering. For the segmentation step, our method utilizes a simple Gaussian distribution model with positional covariances calculated directly on the intensity image. The tracking step is implemented using a classic Kalman Filter that estimates velocity and position of each object in Cartesian coordinates. The sonar detections in the close range of the observable area are compared to the camera detections for validation. Also automatic parameter optimization is used to maximize correlation with camera detections. Additionally, the proposed method is applied to the Caltech Fish Counting Dataset and compared with a deep learning method based on YOLOv5.