Real-time Vision Based Obstacle Detection in Maritime Environments

Automatic obstacle detection is a key feature for unmanned surface vehicles (USV) operating in a fully autonomous manner. While there are currently many approaches to obstacle detection in maritime environments (e.g., LiDAR, radar) the proposed approach resorts to standard, inexpensive RGB cameras to perform the detection of such obstacles. Recent advances in deep neural network detectors achieve state-of-the-art detection results, and some one-stage networks achieve very good results while maintaining inference times small enough to be compatible with real-time capabilities on low-cost embedded processing units.In this paper, we train the YOLO v4 network to detect different types of ships, using publicly available maritime datasets. After training, we evaluate the obtained network on the processing unit located onboard the UAV with respect to detection accuracy and real-time processing capability, thus demonstrating that the presented detection method can be considered a robust, fast, flexible, and inexpensive approach to obstacle detection in USV applications.