无人飞行器辅助的地面无人车辆视觉道路检测方法研究

Vision based road detection is a key technique for unmanned ground vehicles. Monocular vision based method will be the future development trend of road detection due to its low cost, easy installation and low energy consumption. However, the complexity and diversity of road scenes bring great challenges to it. Researchers have proposed a great many methods to tackle those challenges, in which deep learning based methods acting in a supervised manner improve a lot the accuracy of monocular vision based road detection method. Yet, it is not competent in handling unknown environment. In the meantime,limited vision range for a camera mounted on the UGV is obviously present. Thereby, this project is aiming at designing novel road detection method through resorting to a UAV helping a UGV in a self-supervised learning manner. First, we will study road detection related superpixel extraction and similarity measure in order to best consider the property of road scene and the efficiency of our method; After investigating model representations for both road shape prior and road scenery prior, we make use of graph method to extract road areas both in the aerial image and the UGV's image. Lastly, a joint self-supervised learning mechanism for UAV and UGV is studied so as to further improve the adaptation of our method to complex road scenery. The achievements of our project could help boost the safety and intelligence of UGVs and improve the accuracy and ability of adapting to scenes. Moreover, it could also bring some benefits to other research domains in computer vision, such as scene understanding and multi-robot cooperation.

基于视觉的道路检测是无人车(UGV)导航的核心技术。基于单目视觉的道路检测方法因其低成本、易安装和低能耗将成为发展趋势。然而场景的复杂性和多样性给该方法带来巨大挑战，学者们提出很多解决方法，其中深度学习方法大大提高了单目视觉道路检测的准确性，但它对未知环境的适应能力不强。此外，车载摄像机视场受限问题也逐渐突显。针对这些问题，本项目研究一种在线自学习方式下利用无人机(UAV)辅助UGV实现道路检测的方法。具体包括：研究道路检测任务相关的超像素提取和相似性度量方法以充分考虑道路场景的特点以及道路检测方法的效率；研究道路形状先验、道路场景先验的模型化表达，并结合图理论实现准确的航拍图像道路检测和车载图像的道路检测；研究UAV和UGV联合在线学习机制，以进一步提升道路检测的准确性和环境适应能力。本项目的研究有望提升UGV视觉导航的智能化程度，且为场景理解、多机器人协同等方向提供有益的借鉴。

本项目主要研究如何更好地利用无人机来辅助无人车实现鲁棒的视觉道路检测。研究内容主要包括1）适合道路检测任务的超像素提取与相似性度量、2）基于道路场景先验的无监督道路检测方法以及3）无人车和无人机联合在线学习机制。研究得到的重要研究结果有1）提出了基于虚警和漏检对抗机制的目标分割方法；2）提出了基于红外和可见光融合的道路检测方法；3）提出多传感器分层融合的道路检测方法；4）提出无人机俯视视角和无人车前视视角联合学习的道路检测方法；5）提出无人机俯视视角下对合作车辆的单目标跟踪方法；6）提出基于图卷积网络的目标关联方法。此外，本项目以试验验证为目的，构建了无人车视角下同一场景不同季节道路图像数据集、无人机道路检测图像数据集、无人车和无人机协同道路检测数据集、红外与可见光融合道路图像数据集、多地表覆盖道路图像数据集、多场景水体检测图像数据集、小目标分割图像数据集。本课题的研究丰富了多视角、多模态环境感知方法和技术，验证了无人机对无人车环境感知的较好的辅助作用，研究成果正逐步用于军用目标侦察、救灾搜救、无人驾驶等应用中。