Zhijian Huang, Bingwei Hui, Shujin Sun
School of Computer Engineering and Applied Mathematics, Changsha University, Changsha, China.
Hunan Province Key Laboratory of Industrial Internet Technology and Security, Changsha, China.
Front Comput Neurosci. 2022 Jul 15;16:930827. doi: 10.3389/fncom.2022.930827. eCollection 2022.
Training infrared target detection and tracking models based on deep learning requires a large number of infrared sequence images. The cost of acquisition real infrared target sequence images is high, while conventional simulation methods lack authenticity. This paper proposes a novel infrared data simulation method that combines real infrared images and simulated 3D infrared targets. Firstly, it stitches real infrared images into a panoramic image which is used as background. Then, the infrared characteristics of 3D aircraft are simulated on the tail nozzle, skin, and tail flame, which are used as targets. Finally, the background and targets are fused based on Unity3D, where the aircraft trajectory and attitude can be edited freely to generate rich multi-target infrared data. The experimental results show that the simulated image is not only visually similar to the real infrared image but also consistent with the real infrared image in terms of the performance of target detection algorithms. The method can provide training and testing samples for deep learning models for infrared target detection and tracking.
基于深度学习训练红外目标检测与跟踪模型需要大量的红外序列图像。获取真实红外目标序列图像的成本很高，而传统的仿真方法缺乏真实性。本文提出了一种将真实红外图像与模拟的三维红外目标相结合的新型红外数据仿真方法。首先，将真实红外图像拼接成全景图像作为背景。然后，在尾喷口、蒙皮和尾焰上模拟三维飞机的红外特性，将其作为目标。最后，基于Unity3D将背景和目标进行融合，在其中可以自由编辑飞机的轨迹和姿态，以生成丰富的多目标红外数据。实验结果表明，模拟图像不仅在视觉上与真实红外图像相似，而且在目标检测算法性能方面也与真实红外图像一致。该方法可为红外目标检测与跟踪的深度学习模型提供训练和测试样本。
