在无人机平台上部署实时且高效的目标跟踪算法是计算机视觉领域的研究热点之一。论文对孪生跟踪算法中的SiamRPN模型进行了改进，设计了能够在机载处理器上实时运行且与SiamRPN性能接近的单目标跟踪模型Siam-RPN-V3，在NVIDIA Jetson TX2处理器上，SiamRPN-V3将SiamRPN的推理速度由14 FPS(Frames per second)提高到了25 FPS。在伺服控制部分，论文提出了两阶段单目测距算法，使无人机对地面目标跟踪摆脱了高度限制;同时使用四台相机实现了全局视野平台和多相机切换策略，提高了对地面快速移动目标的跟踪能力。
Real-time Moving Object Tracking of UAV Based on On-board Muitl-camera
Deploying real-time and efficient target tracking algorithms on unmanned aerial vehicle(UAV)is one of the re-search hotspots in the field of computer vision.This paper designs a single-target tracking model called SiamRPN-V3,which can run in real-time on the on-board processor and has a performance close to SiamRPN,and on the NVIDIA Jetson TX2,Siam-RPN-V3 improves the inference speed of SiamRPN from 14 FPS to 25 FPS.This paper also proposes a two-stage monocular ranging algorithm,which enables the UAV to track ground targets free from the height limitation.Meanwhile,a global field of view platform and a multi-camera switching strategy are implemented using four cameras to improve the tracking capability of fast-moving ground targets.
国家科技期刊平台
NSTL
万方数据
