Maritime Object Detection for Autonomous Surface Vehicles through Distinct Problem Understanding

Abstract

In recent years, high-performance deep learning-based detection algorithms have been rapidly advancing, raising high expectations for autonomous vehicles, in particular, autonomous navigation. However, current detection studies primarily focus on improving the performance of general-purpose detection. Considering the performance limitations and resource constraints, the pursuit of detecting all maritime objects with the highest performance is not always practical. In this paper, we categorized detection performance into two objectives: (1) safe navigation and (2) surveillance/reconnaissance while describing characteristics of each objective in terms of purpose, priority, sensor weighting, and usage of additional information. Then we proposed a two-stage structure that can effectively handle these objectives. The designed algorithm in this structure selectively applied the 1st stage detection results to the corresponding objects based on the objective, allowing for efficient resource utilization for algorithm execution while still ensuring a minimum level of performance through the continuous operation of the 1st stage detection results. We have also provided example results to show the effectiveness of our proposed method.