碎云环境下GF-4卫星对运动舰船的目标跟踪

A Method for Moving Ship Target Tracking of GF-4 under the Condition of Broken Cloud

针对"高分四号"(GF-4)卫星可进行连续观测和图像分辨率低的特点,文章提出了一种在碎云条件下用GF-4卫星对运动舰船目标进行检测和跟踪的算法。该算法基于区域候选的卷积神经网络(R-CNN)模型构建思路,首先利用双参数恒虚警率(CFAR)完成候选目标区域提取,以充分利用卷积神经网络(CNN)中LeNet网络在识别简单弱小目标时识别速度快的优势;然后对目标进行静态特征提取和鉴别;最后利用地理航行中的中分纬度法与全局最近邻(GNN)算法相结合进行目标关联和跟踪,形成了目标航迹并获取目标运动信息。文章选取GF-4卫星序列图像对所提算法进行试验,并通过舰船自播报(AIS)信息进行验证,结果显示:该算法能在一定碎云条件下排除碎云目标的干扰,有效地进行运动舰船目标检测与跟踪,具备较强的工程应用价值。

Based on GF-4 satellites’ features of continuous observation and low image resolution characteristics, this paper presents a moving ship target detection and tracking algorithm for GF-4 satellite under the condition of broken cloud. This algorithm refers to the construction idea of convolution neural network(R-CNN) model based on region candidate, and uses double parameter constant false alarm rate(CFAR) to complete candidate target region extraction. In order to make full use of the advantage of LeNet network in convolution neural network(CNN) to identify simple and weak targets, static feature extraction and identification are carried out. Finally, the middle latitude method in geography is combined with the global nearest neighbor(GNN) algorithm to coordinate and track the target track and obtain the target motion information. This paper selects the GF-4 sequence image and the ship self-broadcast newspaper(AIS)information to carry on the experiment verification, proves that the algorithm could eliminate the interference of the broken cloud target under certain cloud condition, and effectively carry on the moving ship target detection and tracking. It has strong engineering application value.