Vision-based Object Tracking in Marine Environments using Features from Neural Network Detections

Autonomous decision support is desired to enable navigation with a temporally unattended bridge or to have the vessel navigated remotely. In order to have safe navigation, it is crucial to correctly interpret the current situation given any scenario. Proper perception of the surrounding environment is essential for good situational awareness. This paper suggests a method for tracking objects that have been detected by a neural network. The method utilises features that have been computed during the detection step, thereby ensuring good features that are representative for the given objects while saving the time it would take to compute new features. The suggested method is evaluated on data acquired in Danish near-coastal waters. Evaluation shows that the tracking method is able to track the detections well with few switches of object identity. The method is shown to outperform a similar tracking algorithm, while keeping the speed needed for real-time applications.