Image-based Localization in Dark and Turbid Underwater Environments

This dissertation explores robust localization methods for autonomous underwater vehicles in dark and turbid underwater environments. This dissertation addresses poor visibility, suspended particles, and limited features, which hinder real-time visual perception and localization. It aims to develop innovative techniques that can enhance image quality, estimate motion blur parameters accurately, and improve camera pose estimation, ensuring reliable underwater navigation.

The primary aspects this dissertation considers and the main contributions are as follows.

Underwater Image Restoration and Enhancement: A dual transmittance estimation-based method integrating boundary constraints and local contrast estimation with adaptive ambient light estimation and color correction, is proposed. In addition, a transformer-based feature fusion framework is designed by combining physical model-estimated transmission parameters, which can enhance the visual quality and robustness of underwater images.

Motion Blur Parameter Estimation: A cepstrum morphological feature-based method is developed for estimating motion blur parameters, combining frequency-domain feature extraction and least squares ellipse fitting. This approach can accurately estimate blur angle and scale parameters, improving the performance when handling small-scale blurs typical in underwater imaging.

Transformer-Based Image Enhancement Guided by Physical Blur Estimation: An improved transformer-based image enhancement framework that performs physical blur estimation via Wiener filtering is developed. The explicit incorporation of the point spread function-related information enables precise restoration of image structures and details, outperforming conventional image enhancement techniques.

Monocular Camera Pose Estimation: A monocular pose estimation method based on a parallel perspective error model is developed. This model utilizes a line segment error propagation model to enhance the stability and accuracy of pose estimation, achieving high performance in both simulated and real-world underwater scenarios.

Experiments confirm that the proposed image restoration and enhancement methods outperform existing techniques in subjective visual effects and objective quality metrics. The motion blur estimation method demonstrates high effectiveness in small-scale blur scenarios, and the proposed pose estimation method can improve the localization accuracy of underwater robots. The research findings have significant application potential in underwater robot navigation, environmental monitoring, and underwater exploration. Future research could focus on optimizing computational efficiency, extending the application scenario types, and exploring more efficient sensor fusion strategies.