针对当前军用车辆观瞄系统存在全域感知能力差、危险目标识别率低的问题, 本文提出并实现了一种基于深度学习的智能红外全景识别系统。首先根据可靠性和低功耗原则设计了该系统的硬件部分; 其次, 为解决红外图像中行人和车辆识别率低、实时性差的问题, 设计了一种 BIRD(Brisk InfraRed Detection, BIRD)深度神经网络, 实现对行人或车辆进行检测与识别; 此外, 基于旋转平台多源传感器参数, 实现红外相机快速帧间连续拼接; 实验结果表明, 本系统在实时输出 360.红外全景影像的同时, 能对当前视场中的目标进行同步检测识别, 在保证相同准确率的前提下, 所采用的 BIRD网络与 RCNN网络相比, 处理 PAL视频的平均耗时减少 175 ms, 表现出了实时、稳定的工作性能。

In view of the problems of poor whole-domain perception ability and low recognition rate of dangerous targets in current military vehicle viewing and aiming systems, this paper proposes and implements an intelligent infrared panoramic recognition system based on deep learning. First, the hardware part of the system was designed according to the principle of reliability and low power consumption; second, to solve the problems of low recognition rate of pedestrians or vehicles and poor real-time performance, we designed the BIRD network to detect and identify pedestrians or vehicles; in addition, fast interframe continuous stitching of infrared cameras is realized on the basis of the multi-source sensor parameters of a rotating platform. The experimental results demonstrated that in the case of the simultaneous output of 360° infrared panoramic images, the system could synchronously detect the target in the field of vision. On the premise of ensuring the same accuracy, the system reduced the average time of PAL video processing by 175 ms compared with the RCNN network, thereby confirming the real-time and stable performance of the BIRD network.