低照度环境下基于红外图像的停机坪目标监视系统关键技术研究

Ramp area is an important part of the airport ground. Compared with runway and taxiway, the distribution density and types of moving targets is large in the ramp area. Therefore, the existing airport ground monitoring systems have obvious shortcomings for the ramp area in low-light conditions. In order to enhance monitoring for the ramp area, the project intends to use thermal infrared camera which provides infrared images and is a fixed monitoring equipment in the ramp area, studies moving target detection and recognition, location and tracking, abnormal behavior understanding problems in the ramp area. First, the project explores infrared image saliency detection and feature extraction of typical targets in ramp area, builds infrared image-based mechanism for target detection and recognition. Second, the project studies target location and tracking algorithms, and builds a new framework of target long time tracking algorithm based on infrared images, guided by the idea of tracking-learning-detection and compressive tracking algorithms. The tracking algorithm studied by this project must be robust in the condition of target appearance variations which contain both the intrinsic variations such as pose changes and shape deformation and extrinsic variations such as occlusion. Finally, on the basis of the tracking trajectory(human and vehicle in the ramp area) and semantic feature model of monitored targets(human), combining the airport surface operation rules, the project studies abnormal behavior understanding. In sum, the research of this project can improve the level of the ramp area monitoring and the safety of the ramp area in application, and develop a new method of infrared image-based saliency detection, target recognition and tracking, and activity understanding in theory.

停机坪是机场场面的重要组成部分，相比于滑行道与跑道，停机坪区域内目标分布密度大、类型多，现有的场面监视系统对停机坪区域的监视都存在着明显的不足。为了增强在低照度条件下对停机坪区域的有效监视，本项目拟利用固定安装的热红外摄像机提供的停机坪红外图像，研究低照度环境下停机坪活动目标的检测与识别，定位与跟踪，异常行为检测问题。首先，探索针对红外图像的显著性检测和停机坪典型目标特征提取方法，实现对图像中目标的有效检测与识别；其次，根据识别出的目标，结合跟踪-学习-检测算法和压缩跟踪算法思想，研究停机坪目标的持久跟踪方法，确保目标在遮挡、形变情况下的鲁棒跟踪；最后，以跟踪轨迹（车和人）和目标语义特征模型（人）为基础，结合机场场面运行规则研究停机坪目标异常行为检测。通过本项目的研究，在应用上将提高停机坪监视水平，改善停机坪的安全性，在理论上发展针对红外图像的显著性检测、目标识别、跟踪与行为理解技术。

本项目针对低照度环境下的机场停机坪，利用热红外图像研究场面目标检测识别，定位跟踪与行为理解，以此弥补现有场面监视技术对停机坪监视的不足，提高低照度环境下的停机坪区域安全性。项目历时三年，主要完成如下工作：第一、提出了基于显著性分析的红外图像的目标检测识别新方法，该方法利用红外图像中目标在频域内相较于背景更加显著的特点，通过频域显著性计算得到红外图像的显著图，从而确定目标的粗略位置，然后通过自适应阈值分割显著图，提取出感兴趣区域，进一步在感兴趣区域中计算多尺度窗口的显著度，完成红外目标检测。第二、探索出红外图像中目标持久跟踪新技术，提出了一种的自适应特征融合方法，将多种单一特征进行自适应融合，得到一个判别性更强的融合特征；在此基础上，提出了一种针对红外图像的卷积特征提取方法，将灰度特征和浅层卷积层提取出的特征进行自适应特征融合，可以明显提高跟踪器的跟踪性能，通过和主流跟踪算法的对比，定量定性的分析了该特征的跟踪准确性和鲁棒性。第三、建立基于红外图像的异常行为检测新途径，发展具有红外图像成像特点的图像分析新理论，为了在低光照条件下停机坪场面的视频监控中实时鲁棒地对人体行为进行识别，本项目将两支网络的训练结果分别从可见光数据迁移到红外数据，实现了两种数据介质之间的知识迁移，并从红外数据中学习到了高效的特征；在此基础上提出了一种将三维卷积神经网络结构与双流网络结构相结合的方法，通过对两支网络最后一层卷积层的输出特征做最大值池化、平均值池化、直接相加的方式，在特征层将两支网络相结合，进一步提高了特征的判别力。

内容获取失败，请点击重试