提出并设计了一套应用于小尺寸、高频抖动的微型仿生扑翼飞行器的实时视觉系统, 包括构成硬件基础的微型摄像头模块、5.8GHz传输模块和无线图传接收模块, 以及构成软件模块的两种算法, 其中一种是利用角点检测与LK光流法、基于卡尔曼低通混合滤波的实时电子稳像去抖算法, 另一种是基于YOLOv3的人像识别算法。进行了微型仿生扑翼飞行器的飞行硬件搭载测试实验, 实验结果表明, 所提出的电子稳像算法可以去除x轴66.56%、y轴73.15%的抖动, 人像识别算法可以达到96%以上的准确率, 同时图传分辨率在720×480像素时, 稳像去抖算法时延为6.33ms, 人像识别算法时延为20.5ms, 总时延满足实时要求, 完成了实时去抖与人像识别系统的设计目标。

A real-time vision systemis designed for flapping-wing micro air vehicles (FMAV) withlightweight and high-frequency vibration. The system includes the hardware and software modules. The hardware is consists of a micro camera module, a 5.8GHz transmission module and a wireless image transmission receiving module. And the software modules include two algorithms, one is a real-time video stabilization algorithm based on corner detection, Lucas Kanade optical flow and Kalman low-pass hybrid filtering, and the other one is a person recognition algorithm based on YOLOv3. Flight tests on FMAV with hardware were carried out. The experimental results indicate the video stabilization algorithm can diminish 66.56% vibration of the x-axis and 73.15% vibration of the y-axis. The person recognition algorithm can achieve an accuracy of higher than 96%. The video stabilization algorithm costs 6.33ms and person recognition algorithm costs 20.5ms when the image resolution is 720×480 pixels, thusthe total delay meets the real-time requirements. In conclusion, the design goal of real-time vision system has been achieved.