Jellyfish detection based on K-FOE residual map and ring segmentation

Target detection from underwater videos is a hot and difficult area in computer vision, especially when the camera has ego-motion. A jellyfish detection system is proposed for processing video streams captured by moving cameras mounted on remotely operated vehicles (ROVs). The background motion vector convergence point, or focus of expansion (FOE) is first found by solving equations with optical flows and then modified by Kalman filter prediction (K-FOE). Object templates are initialized by binarizing the K-FOE residual map. Subsequently, a ring segmentation subsystem is used to update the primitive object mask according to the distance between the object mask center and the K-FOE. All objects are extracted after their updated masks are obtained. Experimental results on real video data show that the proposed system can not only reduce fault detection but also extract small size jellyfish objects well.