机场场面监视中复杂目标检测与实时跟踪关键深度学习技术研究

With the rapid development of the national economy, the volume of air transport in China has maintained a rapid growth for a long time. In order to improve the security in airports, it is necessary to analyze the surveillance contents intelligently under the condition of the existing airport scene monitoring. This project is devoted to the in-depth study of robust object detection and real-time online multi-object tracking under the airport scene. The relevant deep learning techniques will be utilized and improved for the high precision and high efficiency of small object detection with ultra high resolution, which lays a foundation for research in intelligent scheduling of the airport and the accident early warning. At the same time, the fine-grained classification on special targets such as aircraft and vehicles will be studied in this project. In addition to high resolution and a wide variety of object categories, the airport scene has to face the interference of weather and light conditions, such as haze and heavy rain and blizzard. These conditions affect the clarity of the video surveillance and cause serious interference to the object detection and tracking. Therefore, this project will also use the deep generative method to study the image enhancement for these special situations to improve the overall performance. The research of this project can fill the gap in this field in China and speed up the construction of the intelligent security airport.

随着国民经济的快速发展，我国航空运输量长期保持高速增长。为了提高机场运作的安全性，需要在原有机场场面监视的条件下进一步对监视内容进行智能化分析。本项目针对机场场面下的多类目标进行检测和跟踪进行深入研究，利用并改进目前基于深度学习的算法，提高其在超高分辨率和小物体检测上的效率和精确度，为机场智能化调度，事故预警打下研究基础。同时在此基础上我们也会对飞机、车辆等特殊目标进行细粒度分类的研究以进一步增加机场运作的安全性，提高机场运行效率。机场场面除了分辨率高，物体繁多之外，还要面临着天气和光线状况的干扰，如雾霾和雨雪天气。这些情况会影响监视视频的清晰度，对目标识别造成严重的干扰。所以本项目也会利用深度生成网络模型针对这些特殊情况进行图像增强的研究，以提高整体的性能。该项目的研究可以填补国内在这一领域的空白，加速智能化安全化机场的建设。

该项目由清华大学深圳国际研究生院和中国民用航空总局第二研究所共同完成，针对机场场面监视中复杂目标检测与实时跟踪，利用并改进目前基于深度学习的算法，提高了其效率和精确度，为机场智能化调度，事故预警打下研究基础。项目的主要内容包含机场场面物体检测、识别和跟踪为一体的整体架构设计，小物体检测与跟踪，机场环境下不同人员的服装分类和属性识别，人员服装属性和车辆的细粒度识别，多目标的实时跟踪，以及针对复杂场景的实例分割。在项目实施过程中发表了21篇高水平学术论文，7项专利，培养博士和硕士共4名，对机场场面的智能化分析积累了实际的验证数据和科研成果转化经验，对今后的进一步研究以及成果落地工作起到了重要的基础作用。