Computer Vision Workflow

CV Workflow

Intro to Computer Vision

By now, everyone should know that we are exploring brush management as an early application of Arrow’s multi-purpose drone. Tied in with this task is the processing of visual signals. That description is intentionally vague, as there are a lot of ways to approach this goal. In this short memo, I will outline just a few possible workflows as we try to test this in the market.

Relevance to Arrow

How does this relate to Arrow’s work? Mostly, it doesn’t. A lot of this R&D is likely to fall squarely on the DevCo side of things. However, it may offer two pieces of insight:

  1. Devices and computation required (or not) to accomplish this task
  2. Future required capabilities anticipated for Project Quiver

1) Just Look

By far, the simplest way to accomplish brush management is visually. The drone operator is tasked with flying the drone, looking for the target plants on a display, and then push a button. A simple, manual process.

Required Materials:

  • Downwards-facing camera
  • VRX video display
  • Remote-controlled dispenser

Limitations:

  • Demands full attention of a pilot
  • Additional time of pilot
  • Requires high-fidelity cameras to judge target plants
  • Risk of duplicate treatments

2) Brute-Force Search

Full grid search of a bounded area

Our second option is where we start to incorporate AI. We can bound a region to explore and trigger a dispensing event for each time out real-time computer vision setup detects a target plant directly under our drone.

Basic process:

  • Outline boundaries of region to explore
  • “Zigzag” or grid explore the full region
    • If the CV tool detects a target plant, dispense herbicide
  • Continue until end of route

Required Materials:

  • Downwards-facing camera
  • Remote-controlled dispenser
  • Onboard compute for CV
  • A trained model to ID target plants in near real-time
  • (Nice to have) VRX video display

Limitations:

  • Requires additional compute onboard
  • Demands well-trained, performant CV model
  • Lots of redundant travel (!)
    • Undermines speed & endurance requirements
  • Risk of duplicate treatments

Commentary

This is called a “brute force” search for a reason. While it does offload much of the decision burden to the onboard companion computer, this technique performs an exhaustive search of the area of interest.

In the example image above the full flight plan covers a distance of 2.81 km. At a rate of 5m/s, completing this task will take about 9 minutes and 20 seconds. We know that time and endurance of our drone are of high importance to end users, so this is unlikely to be a suitable solution for larger operations.

3) 2-Pass Search + Address

The current most interesting solution involves a 2-pass method. Pass 1 uses a higher-flying scouting drone to take images of the region of interest.

The goal of this image collection is to create a comprehensive “map” of the region, then use a CV model on an off-board computer to find the target plants. These targets are assigned waypoints and events, which are then sent to the multipurpose drone, which then performs the treatment in the second pass.

Pass 1: survey images to map

Once here, run your CV to determine which of the features of this map represent target plants. If the map is geo-referenced, it should be fairly straightforward to convert these points to waypoints for a mission planner. Once the mission is decided, send it to the multipurpose drone.

Pass 2: directly addressing the plants

Required Materials:

  • A second drone specifically for scouting
  • Downwards-facing camera
  • Remote-controlled dispenser
  • Software to develop a contiguous map of area of interest
  • A trained model to ID target plants & offboard computer
  • (Nice to have) VRX video display

Limitations:

  • Requires additional drone
  • Robust data handoff: scout drone → computer → Quiver
  • Demanding, complex tech stack

Commentary

Of the three, this is where I picture the future of this workflow going. This not only offloads much of the burden from the drone operator, but is also more efficient by using specialized drones.

In this example, by eliminating the brute force search step for Quiver, its overall mission distance was reduced from 2.81km to 0.82km, dramatically increasing the efficiency of the mission. This offers a promising challenge to the dominant helicopter IPT status quo.

However, there is quite a distance between this vision and a working model. In the meantime, we should continue to enhance the endurance of Project Quiver as we begin to see more possibilities for long-range missions.

Conclusion

TLDR; We’ve got a plan for getting this guy to work with computer vision. Most of this plan will be tested and incorporated in satellite devices + software around Project Quiver.

If engineering takes one thing away from this:
For this to work, Project Quiver needs to be a robust action drone with an emphasis on endurance. If we can count on Quiver to be a reliable platform, we can unlock a lot of interesting pos

3 Likes

Nice postulation.

On option 3, another possible take would be to just have 1 drone with 2 batteries. One in, another spare.

If the batteries are easy swap, i.e pullout, push in, fasten, go, then down time is minimised to seconds, and the user only needs to deploy one drone.

So first pass generates the waypoint map, then second pass does the work. Reduced costs overall

3 Likes

I prefer the method 3 .

Because pre-scout geographic survey data can provide very good result and it’s a mature technology. It can create contiguous image of large areas in a short period of time for counting standalone vegetation objects. It can also benefit the project by creating various maps and historical records for working area. But it need to take several hours to generate the image from aerial photos.
(Also I was a supervisor of that so you can ask me everything about that…)

For the Software to develop a contiguous map of area of interest (scientifically called Digital Orthophoto map or DOM), We can use:

  • Open Drone Map (Best in free selections, hard to deploy and use)
  • iTwin Capture (Most robust and accurate, most expensive, Made in US)
  • Metashape (Maximized survey and production functions)

Then we can view the DOM by normal photo viewer, and handling using one of these software below to mark individual plants and create the mission way points manually (It will include real-world coordinates) :

  • QGIS (Free)
  • Global Mapper (User friendly)
  • ArcGIS Pro (Top tier, not expensive, Made in US)

For the machine vision and automatic plant recognition from DOM, I’m very poor at programming but I hope this can help. It’s very rich.

FIELDimageR - OpenDroneMap™


(Discord) This is a DOM made from full frame sensor and iTwin Capture with regular mid level workflow. To experience the quality, You can try identify individual plants, vehicles, or even personals in full-screen.

3 Likes