Weather-aware object detection method for maritime surveillance systems

Abstract

The development of machine learning-based maritime object detection technology aims to assist ship operators in maritime surveillance. However, as maritime environments can be quite complex, developing object detection models that can handle these situations is a challenging research problem, particularly when dealing with adverse weather conditions like rain and haze. While prior research has attempted to remove weather noise and improve object detection models under various weather conditions, they are limited by computing resources and hard to adapt to the constantly changing weather conditions of maritime environments. Preventing performance degradation as weather conditions shift is a significant challenge in maritime surveillance systems. To overcome these challenges, this paper proposes a weather-aware object detection method, Weather-OD, that employs an on-board edge and on-shore cloud-based system for maritime surveillance. It employs specialized machine learning models for object detection, which can be dynamically selected based on the weather conditions to ensure highly accurate object detection with low latency at sea. Weather-OD continuously improves accuracy by periodically training the models with newly collected datasets during voyages, and efficiently manages the life cycle of multiple object detection models, taking into account the constraints of limited edge computing resources. In addition, Weather-OD uses synthetic image data with weather noise to supplement the training data under different weather conditions. We conducted an evaluation of our weather-aware object detection models using a maritime benchmark dataset, the Singapore Maritime Dataset. Our experimental results demonstrated the feasibility of our mechanism with weather classification and a significant improvement in the mean Average Precision (mAP) of maritime object detection in rainy and hazy conditions. Additionally, our approach enables the continuous improvement of object detection accuracy through model retraining with small datasets.