火箭图像目标识别与跟踪是图像目标识别的重要应用领域，是实现火箭测试发射、飞行控制的重要支撑，对火箭目标跟踪、姿态分析控制具有重要意义。上升段的火箭目标视频图像跟踪是火箭飞行测控的重要阶段，但目前对火箭上升段的视频图像跟踪主要依靠人工手动操作云台控制器，图像跟踪存在跟踪滞后、画面抖动等问题，跟踪效果受人为因素影响较大。结合全卷积理论和深度学习方法，提出一种基于全卷积深度神经网络的火箭图像目标识别与跟踪方法，采集火箭发射及上升段的图像作为样本，构建、训练全卷积网络模型，采用端到端的语义分割方法，在深度分类网络的基础上，实现火箭目标在像素级别上的语义判断，具有较好的识别率和鲁棒性。在火箭目标识别的基础上建立云台控制模型，通过对云台的智能控制获得火箭上升段的高质量图像，完成对火箭目标的跟踪。

Rocket image target recognition and tracking is an important application field of image target recognition，which is an important support for rocket test launch and flight control，and has great significance for rocket target tracking and attitude analysis and control.Image tracking of rocket target in ascending stage is an important stage of rocket flight measurement and control，but at present，the video image tracking of rocket in ascending stage mainly relies on manual operation of pinion controller to achieve rocket tracking in the image.The tracking image has some phenomena such as tracking lag and picture shaking，and the tracking effect is greatly affected by human factors.Combining full convolution theory with deep learning method，a method of rocket image target recognition and tracking based on full convolution deep neural network was proposed.Images of rocket launch and ascending flight were collected as samples，and a full convolutional network model was constructed and trained.An end-to-end semantic segmentation method was adopted to realize semantic judgment of rocket targets at pixel level on the basis of deep classification network，with good recognition rate and robustness.Based on the recognition of the rocket target，the PTZ control model was established，the high-quality image of the rocket ascent stage was obtained through the intelligent control of the PTZ，and the tracking of the rocket target was realized.