Artificial Intelligence for Vision Based Navigation

传统的地理定位技术（如GPS）在某些环境（如室内）信号弱或精度降低，或者在某些严格的情况下被拒绝。基于视觉的信息不仅易于使用简单和低成本的设备捕获，而且具有强大的模式识别和人工智能方法武装的语义感知能力。因此，如何利用基于视觉的信息在上述情况下进行定位和导航是一个具有挑战性的研究课题，但对许多实际应用具有潜在的高影响。这个博士项目的目标是使用基于视觉的信息，通过人工智能方法，特别是深度学习技术进行定位和导航。近年来，深度学习技术取得了惊人的进步，并被应用于许多具有挑战性的领域，如自然语言处理和计算机视觉。由于其广泛的代表性，深度学习也应该在基于视觉的导航应用中发挥关键作用。该博士项目的主要工作方向包括-基于深度学习的图像匹配技术，用于使用现有的地形测量地图或低分辨率卫星图像进行地理定位（即基于绝对图像的导航）；还将应用与迁移学习有关的技术来缩小预注册地图（例如，地形测量地图或低分辨率卫星图像）与真实世界RGB图像之间的差距。该方法涉及的技术包括地标检测、图像匹配和位置估计。许多深度学习方法可以应用于这些任务，例如用于地标定位的目标检测，用于图像匹配的迁移学习，以及用于位置估计的一些决策算法。-基于深度学习的光流技术，用于自我运动（即基于相对图像的导航）估计，这是使用视频数据进行导航的关键组成部分。这种方法也提出了一些理论和实践上的挑战，如数据关联、遮挡以及在利用单目相机时缺乏直接度量信息。使用深度学习算法估计光流也是该项目的一个有趣的子主题。- 3D环境重建和语义理解与额外的传感器信息（例如，RGB-D），以协助精确定位和导航。例如，重建三维环境和检测三维物体有助于在导航过程中避免潜在的障碍物。这种方法与同时定位和映射（SLAM）密切相关，其中在构建未知环境的地图的同时保持对代理位置的跟踪。

Traditional geolocalisation techniques (e.g., GPS) suffer weak signal or precision decrease in some environments, e.g., indoor, or are denied in certain strict situations. Vision based information is easy to capture only with a simple and low-cost device, but also with strong pattern recognition and semantic sensing power armed by artificial intelligence methods. Therefore, how to utilize vision-based information to localize and navigate in the above situations is a challenging research topic but potentially has high impacts on many real-world applications.The goal of this Ph.D. project is to use vision-based information for localization and navigation with artificial intelligence methods, specifically, deep learning technique. In recent years, deep learning technology has achieved stunning progress and has been applied to many challenging fields, e.g., natural language processing and computer vision. Thanks to its generalized representative power, deep learning should also play a critical role in vision based navigation applications. The main work directions in this Ph.D. project include- Deep learning based image matching technique for geolocalisation (i.e., absolute image based navigation) using existing ordnance survey maps or low-resolution satellite imagery; Transfer learning related techniques will be also applied to bring the gap between pre-registered maps (e.g., ordnance survey maps or low-resolution satellite imagery) and real-world RGB images. Techniques involved in this approach include landmark detection, image matching and position estimation. A number of deep learning methods are able to be applied in these tasks, i.e., object detection for landmark localization, transfer learning for image matching, and some decision making algorithms in position estimation.- Deep learning based optical flow technique for egomotion (i.e., relative image based navigation) estimation, which is a critical component for navigation using video data. This approach also presents several theoretical and practical challenges, such as data association, occlusions, and lack of direct metric information when exploiting monocular cameras. The estimation of optical flow with deep learning algorithms is also an interesting subtopic of the project.- 3D environment reconstruction and semantic understanding with additional sensor information (e.g., RGB-D) to assist in precise localisation and navigation. For example, to reconstruct 3D environment and detect 3D objects is helpful to avoid potential obstacles during navigation. This approach is closely related to Simultaneous Localization And Mapping (SLAM), in which a map of an unknown environment is constructed while simultaneously keeping the track of an agent's location.