Skip to main content

Notes from Drone research March 21 2025

·506 words·3 mins·
Author
Mark Ogata
AI and Robotics Researcher at BAIR

Todo
#

Overall plan (feel free to edit this, I just wrote this down in 1 min)
#

  1. Get raspi and camera
  2. Get communication between voxl2 and raspi working
  3. Get two camera streams from CSI cameras to displaying in raspberry pi
  4. try out algorithms for moving object or drone detection (can test by just walking around with the setup and moving other objects in the room simultaniously)
  5. mount the raspberry pi and cameras in the drone
  6. Adapt the algorithm to work on omnidirectional on the drone
  7. Test detection algorithm outdoors by looking at detections while having the drone hover and walking around, throwing objects, flying drones
  8. get prioritization algorithm and tracking algorithm working
  9. Test it indoors
  10. Test it outdoors

hardware
#

360 camera with raspberry pi 5 (dont use 4 because we need 2 camera slots)

Setups to model in Onshape:
#

FOV analysis above and below right and left front back Once FOV looks good, figure out mounting mechanism and compute weight:

  • < 500g (add up the weights of everything raspi, cameras, mounts, screws, power mechanism)

software
#

  • Literature search:
    • Problem statement: there exist 2 180 images, what detection algorithm do we run to detect flying objects
      • There are SOTA algorihtms according to Zakhor: try google scholar search on “moving object detection” + sky
      • This paper from Zakhor could be related: https://www-video.eecs.berkeley.edu/papers/lyang/Drone_Object_Detection_Using_RGB_IR_Fusion.pdf
      • look at related work when searching for papers and get top mentioned papers
      • idea from Zakhor: use a simulator like unreal engine to create synthetic data and train a neural network. (I think this is overkill and much simpler methods would be more reliable and fast but idk)
  • How do we decide between multiple tracked objects?
  • How to re-identify lost objects and occluded objects
    • there are algorithms for this according to Zakhor. Need literature search on this

Definition of a threat (discussed with Zakhor): Not a ground object and moving
#

on the debate: USB cameras -> voxl2 vs CSI cameras -> raspberry pi with classical vision -> voxl2
#

Bandwidth limitations of USB means we would most likely encounter problems if hooking up two cameras via USB. Comparison of using raspi vs usb camera

  • Pros
    • fast CSI connection from cameras mean higher frame rates and resolutions possible (probably leads to better detection and tracking)
    • Raspberry pi is very common platform so many cameras that do >180 fov
  • Cons
    • Raspberry pi has low compute compared to Voxl so we have limited ability to run big algorithms (although there are things like google coral that can boost compute)
    • Need to find a way to power the raspberry pi
    • extra complexity of raspberry pi to voxl2 communication (although this should be very common problem and easily solved)

Notes
#

Links to hardware (usb cameras too if thats useful)

https://www.arducam.com/product/arducam-imx291-usb-camera-with-case-b026101/

https://www.webcamerausb.com/elp-fisheye-wide-view-angle-180degree-2mp-full-hd-1080p-50fps-usb-30-camera-module-sony-imx291-uvc-plug-play-driverless-otg-webcam-p-263.html

Mathematically proved >110 degree fov cameras can do full coverage with 6 cameras

anything less than 180 needs 4 cameras

180 degrees needs 2 cameras (this one is the best)

Related

Cool Startups and Labs
·293 words·2 mins
Flow Matching - How Image Generation Works
·1064 words·5 mins