Automated guided vehicles (AGVs) are Internet of Things robots that navigate automatically as guided by a central control platform with distributed intelligence. Different methodologies have been proposed for AGV visual tracking applications. However, vision-based tracking in AGVs usually confronts the problem of time delay caused by the complexity of image processing algorithms. To balance the trade-off among algorithm complexity, hardware cost and performance, precision and robustness are usually compromised in practical deployment. This paper proposes a prototype design of a visual tracking system. Edge computing is implemented which migrates computation intensive image processing to a local computer. The Raspberry Pi-based AGV captures the real-time image through the camera, sends the images to the computer and receives the processing results through the WiFi link. An improved Camshift algorithm is developed and implemented. Based on this algorithm, the AGV can make convergent prediction of the pixels in the target area after the first detection of the object. Relative coordinates of the target can be located more accurately in less time. As tested in the experiments, the system architecture and new algorithm lead to reduced hardware cost, less time delay, improved robustness and higher accuracy in tracking.

自动导引车 (AGV) 是物联网机器人，可在具有分布式智能的中央控制平台的引导下自动导航。已经为 AGV 视觉跟踪应用提出了不同的方法。然而，AGV中基于视觉的跟踪通常面临图像处理算法复杂性导致的时间延迟问题。为了平衡算法复杂度、硬件成本和性能之间的权衡，在实际部署中通常会牺牲精度和鲁棒性。本文提出了一种视觉跟踪系统的原型设计。实施边缘计算，将计算密集型图像处理迁移到本地计算机。基于树莓派的AGV通过摄像头实时捕捉图像，将图像发送到计算机并通过WiFi链接接收处理结果。开发并实施了改进的 Camshift 算法。基于该算法，AGV可以在第一次检测到物体后，对目标区域内的像素进行收敛预测。可以在更短的时间内更准确地定位目标的相对坐标。经实验验证，该系统架构和新算法降低了硬件成本，减少了时延，提高了鲁棒性，并提高了跟踪精度。