Visual perception for long-distance and small target detection in autonomous maritime navigation
Ruolan Zhang, Xingchen Ji, Sean Loughney, Jin Wang, Zaili Yang
Abstract
In the pursuit of advancing autonomous maritime navigation, this study aimed to develop a novel architecture designed to enhance the detection accuracy of distant and small targets under the constraints of real-time performance and robustness. Through the innovative integration of the Convolutional Block Attention Module (CBAM) into the detection model's backbone, the study achieved superior feature extraction capabilities tailored for the complexities of maritime environments. Further optimization of the Spatial Pyramid Pooling (SPP) module ensured model compactness and computational efficiency, vital for deployment on edge devices. A key methodological novelty lay in the incorporation of the S-IoU loss function, which offers superior bounding box regression accuracy over the traditional Generalized Intersection over Union, directly contributing to more precise navigation and effective obstacle avoidance. The proposed enhancements collectively yielded a 5.1 % increase in mAP@50 %, accompanied by an 11.2 % reduction in model parameters and a 12.6 % decrease in computational complexity (GFLOPs). These findings underscore the potential of the presented architecture to significantly contribute to maritime safety, presenting an optimized solution for collision avoidance and navigation assistance in congested sea routes and adverse weather conditions.