Detection and tracking of crossing vessels for small autonomous vessels equipped with stereo camera

When Small Autonomous Vessels perform a mission in a non-segregated area, that is, an area where other vessels (manned or not) will also be present, they need to be able to comply with navigation rules, such as collision avoidance. In order to decide to give-way or stand-on, the autonomous vessel must keep track of other vessels. In this work, a vessels detection and tracking solution is presented. The system uses a depth camera, IMU, magnetometer and GNSS as main sensors for this task. The boat detection is performed by the SSD-MobileNet V2 Neural Network pre-trained with the COCO dataset. The tracking is performed by a simple and fast algorithm, and the samples are later filtered using a Kalman Filter. The solution is implemented in the Robot Operating System (ROS) framework. Practical tests were performed, where stationary and moving targets were tracked. Results show that the method performs well and is able to achieve even better results after some fine tuning.