UKHO Celestial Navigation Project

研究问题：该项目将利用夜空图像来预测相机的位置。该项目的关键研究问题包括，机器学习系统是否可以使用夜空图像来回归相机坐标和参数。该项目还将研究允许特征提取和泛化的基本方法，以便充分训练可靠和准确的回归模型。还将探索噪音对系统的影响，评估当相机可能看不到某些物体或引入光污染时的影响。使用的方法：该项目将使用一种无监督的方法来识别图像中的关键天体，从图像中提取关键点以提供对夜空的观测。将使用传统的图像处理和机器学习方法来检测关键点。然后，这些关键点将用于通过在标记的样本上训练模型来回归相机姿势参数。在给定光污染和遮挡等参数的情况下，还将探讨模型对噪声的稳健性评估。这将被用来为用户产生可解释的预测表示，通过在直观的可视化中观察每个身体对预测的重要性。新颖的内容：除了相机回归能力的提高外，项目的新颖性来自应用领域，并使用以人为中心的方法来实现可解释的预测可视化。在开发利用机器学习的导航系统时，对结果的解释是关键，从而允许领域专家仔细审查预测。对图像中恒星遮挡的稳健性方法的发展也是一个关键贡献。

The research questions:The project will look to utilise imagery of the night sky to perform prediction of the cameras position. The key research questions for the project include whether imagery of the night sky can be used by a machine learning system to regress camera coordinates and parameters. The project will also look into underlying methods which allow feature extraction and generalisation in order to adequately train a reliable and accurate regression model. Exploration into the impact of noise into the system will also be explored, evaluating the impact when certain bodies may not be visible to the camera or when light pollution is introduced. The approaches used:The project will utilise an unsupervised approach to identifying the key celestial bodies in an image, extracting the keypoints from an image to provide an observation into the night sky. Conventional image processing and machine learning methods will be used to detect the keypoints. These keypoints will then be used to regress the camera pose parameters by training a model on labelled samples. Evaluation of model robustness to noise will also be explored, given parameters such as light pollution and occlusion. This will then be used to produce interpretable representations of the prediction for the user, by observing the importance each body contributes to the prediction in an intuitive visualisation. Novel content:Besides the improvement in camera regression ability, the project novelty comes from the application domain, and the use of a human-centred approach to interpretable prediction visualisation. The interpretation of the results is key when developing a navigational system that utilises machine learning, allowing for predictions to be scrutinised by the domain expert. The development of methods robust to occlusion of stars in the image is also a key contribution.