Developing a deep scene parser for UAV-acquired images

Sightline是视觉分析系统开发和集成领域的领导者，其核心VtiS（视觉、遥测、智能系统）平台旨在扩大无人机与国防、搜索和救援以及企业智能的态势感知。 Taylor博士是深度学习领域的专家：使用多层处理从数据中自动提取有用的表示和抽象的方法。 我们一起提出了一个场景解析器的无人机采集的图像数据的发展。场景解析是将每个像素分配给一个类别，例如道路、建筑物、人和汽车。它使人类观察者和机器能够识别场景中存在的对象，推理它们的关系，并开发适当的响应。场景解析器将为VtiS平台提供有价值的分析组件，支持Sightline在无人机系统市场的增长。为了处理大量的数据，我们将利用高性能计算的进步，特别是图形处理单元（GPU）。场景解析中的主要算法挑战是标记像素所需的信息通常来自遥远的像素及其标签。深度学习提出了一种整合本地和全局信息的方法。它在非空中场景解析方面取得了一定的成功，例如标记街道场景。然而，这些工作都没有着眼于航空图像，这涉及到额外的研究挑战，如集成多种传感器类型，并利用捕获的图像之间的时间依赖性。我们希望我们的研究能够立即应用于搜索和救援领域。从长远来看，我们预计这项研究的成果也将影响精准农业。

Sightline is a leader in visual analytics systems development and integration - its core VtiS (Vision, Telemetry, Intelligence System) platform is designed to broaden situational awareness across UAV & defence, search and rescue, and enterprise intelligence. Dr. Taylor is an expert in the area of Deep Learning: methods that automatically extract useful representations and abstractions from data using multiple layers of processing. Together we propose the development of a scene parser for UAV-acquired image data. Scene parsing is the assignment of each pixel to a category, for example, roads, buildings, people and cars. It enables both human observers and machines to identify what objects exist in a scene, reason about their relationships, and develop appropriate responses. The scene parser will provide a valuable analytical component to the VtiS platform supporting Sightline's growth in the Unmanned Aircraft Systems market. To tackle the sheer amount of data to be processed, we will leverage advances in high-performance computing, specifically Graphics Processing Units (GPUs). The major algorithmic challenge in scene parsing is that the information necessary for labeling a pixel often comes from distant pixels as well as their labels. Deep Learning proposes a way to integrate both local and global information. It has had some moderate successes in non-aerial scene parsing, for example, labeling street scenes. However, none of this work has looked at aerial images which involve additional research challenges such as integration of multiple sensor types, and exploiting the temporal dependence among captured images. We expect our research to lead to immediate applications in the area of search and rescue. In the longer term, we expect the outcomes of this research to also impact precision agriculture.