AI For Understanding Multi-domain Point Clouds

点云是感知3D数据最重要的数据表示之一，是地图、数字工程、数字孪生、城市分析、机器人、自动驾驶和混合现实等领域的基础应用。摄影测量、激光雷达等数据采集方法和设备已经达到成熟的状态。因此，点云数据可以广泛获得，或者可以以中等成本获得。与此形成对比的是，相对缺乏能够以高可靠性或准确性自动理解和分类3D感测数据的算法。在过去的十年里，深度学习推动了二维图像理解的进步，可以与人类的表现相媲美。尽管在2D方面取得了这些进步，但自动3D感知数据理解的技术，如点云，相对来说还不成熟。基于深度学习的下一代人工智能有望成为自动化3D点云理解的游戏规则改变者。2.3目标和目标y1：使用可用的国家地图点云数据建立多领域，多传感器或多国基线数据集，以基准测试下一代深度学习架构。Y2：使用移动地图系统从OS （Street Drone）和UCL （Robin）获取多域点云数据集，并可能进一步获取设备(例如；手持SLAM扫描仪)。Y3：基于多域点云解译的新型深度学习框架的开发与文档化。对DL架构进行基准测试，并将其与当前的技术状态进行比较。2.3.1范围内：使用机器学习对来自各种来源的3D点云进行分类。ml3.2对点云分类的重新标记方法的研究。2.3.2范围外：从点云创建多边形/多面体特征几何

Point clouds are one of the most significant data representations for sensed 3D data and underpin applications in mapping, digital engineering, digital twins, urban analytics, robotics, autonomous driving, and mixed reality. The data acquisition methods and devices, such as photogrammetry and LIDAR, have reached a mature state. Therefore, point cloud data is widely available or can be acquired at moderate cost. This is contrasted by a relative lack of algorithms that can automatically understand and classify 3D sensed data with high reliability or accuracy. Over the past decade, deep learning has driven progress in 2D image understanding to rival human performance. Despite these advancements in 2D, techniques for automatic 3D sensed data understanding, such as point clouds, are comparatively immature. Next-generation AI based on Deep Learning promises to be a game changer in automated 3D point cloud understanding.2.3Aims and objectivesY1: Using available national mapping point cloud data establish multi-domain, multi-sensor or multi-country baseline dataset to benchmark next-gen DL architectures.Y2: Acquire multi-domain point cloud dataset using mobile mapping systems from OS (Street Drone) and UCL (Robin) and potentially further acquisition devices(e.g. handheld SLAM scanners). Y3: Development and Documentation of novel DL framework for multi-domain point cloud interpretation. Benchmarking and comparing the DL architecture with the current state of art.Y4: Thesis completion.2.3.1In scope:Classifying 3D point clouds from various sources using machine learningInvestigation of relabelling approaches to point cloud classification by ML2.3.2Out of Scope:Creation of polygon/polyherdral feature geometries from pointclouds