A maritime panoramic visual perception framework with joint video stitching and target detection to enhance autonomous surface vehicles navigation

Accurate panoramic visual perception is essential for reliable navigation of Autonomous Surface Vehicles (ASVs). However, existing vision-based methods are constrained by limited fields of view and often suffer performance degradation in complex maritime environments characterised by occlusion, strong light reflections, and resolution loss under real-time constraints. To address these challenges, this study proposes a maritime panoramic visual perception framework that integrates panoramic video stitching and target detection to enable robust 360° perception for ASV navigation. In the stitching stage, cylindrical projection and hash mapping techniques are employed to achieve low-distortion and real-time panoramic video stitching. In the detection stage, a panoramic vision-oriented detection model is incorporated, in which a Context-Gaussian Hybrid Module and a Feature Modulation and Upsampling unit are devised to enhance the detection capability for maritime targets in panoramic images. Experimental results on the Pohang dataset show that the stitching method produces seamless panoramic videos at a resolution of 1500 × 328 with a speed of 20 frames/s. The detection model outperforms state-of-the-art approaches, achieving 91.3% mAP 0.5 and 64.8% mAP 0.5:0.95 at a speed of 84 frames/s. Furthermore, robustness is verified under eight maritime-specific synthetic corruption scenarios. Overall, the proposed framework meets the engineering application requirements of ASV navigation systems. • A joint video stitching and target detection framework for maritime panoramic visual perception is proposed. • A panoramic video stitching method for shipborne multi-camera systems is proposed with cylindrical projection and hash mapping. • A Context-Gaussian Hybrid Module is designed to enhance detection under occlusion and complex backgrounds. • A Feature Modulation and Upsampling unit is introduced to mitigate detection degradation caused by resolution loss. • The proposed framework achieves superior real-time performance and robustness in panoramic perception.