Skydio 是 MIT 校友公司研发的纯视觉自主航拍无人机。最新产品 Skydio 2

Skydio 2

从技术角度上看，Skydio 2 最大亮点是 Perception 和 Motion Planning。

Skydio 2 感知方案

  • 广角相机
    • 6x cameras in trinocular configuration top and bottom
      • FOV 200°
      • COVERAGE 360°
    • 相机标定 Camera Calibration
  • NVIDIA Tegra TX2
    • DNN
      • 9 custom deep networks used in flight
        • People and motor vehicles

感知模块

  • Tracking
    • 光流法？
  • 3D Mapping
    • GPU双目稠密重建？
  • DNN for Detection

How has Skydio’s autonomy gotten better?

The starting point is having much more data coming in and a faster computer, which opens up some new possibilities, but even with all the hardware stuff, the biggest progress is still on the software side. One of the big things that we’ve done is come up with an entirely new depth perception algorithm based on deep learning rather than more conventional [photogrammetric] stereo techniques, which generally work in simple scenes but fail when there’s a lot of repeated structure or textureless surfaces.

What people do is combine contextual scene understanding with photometric information—when we look at a scene, we know what everything is and generally how the world is structured around us. And the combination of that contextual information with the photometric information means that people and animals essentially have perfect depth perception based just on vision. We’ve come up with a deep learning algorithm that mimics this, such that the drone not only uses photometric information but also contextual information. This drastically improves performance in every scenario that we care about, and overcomes a lot of the really tricky cases. For example, flying around in a big textureless room, we see near perfect performance. And with power lines, if it can see any part of the power line, it does the same thing that a person does and assumes that the line extends. It’s a really, really powerful thing.

用直觉来感受一下

参考文献