4 ms·

Sorry, I don't think I can give you a lot of details. Let's say that one of my previous employers did a project which was supposed to be indoor quadcopter for c

by fest 10y ago

Sorry, I don't think I can give you a lot of details. Let's say that one of my previous employers did a project which was supposed to be indoor quadcopter for consumers. Regarding PX4- completely agree. People working on it definitely have high standards for software engineering practices and sense of responsibility. ArduPilot guys are also doing very good job, especially regarding state estimation (sensor fusion). Indoor navigation is sadly not fully solved problem, at least for open-source solutions. I have not tried it, but IMO the best option at the moment seems to be Qualcomm's drone platform, as that has high-resolution, wide angle camera and enough processing oomph for image processing. PX4Flow did work on some surfaces, but did not work too well on highly repetitive textures (e.g. office carpets). The image sensor, although is very good (global shutter, large, sensitive pixels) had limited resolution and even more limited was processor doing image processing. The fusion of flow data was also quite unstable for ArduCopter and PX4 (they both shared the state estimation code at that point). Some of the problems were implementation issues (e.g. a few bad measurements caused the rest of (now valid) measurements to be ignored) but some were fundamental ones. The algorithms in use relied on constant distance between camera and objects on scene (required for calculating velocity in uniform units) which was unrealistic assumption indoors (e.g. flying near walls, over tables, etc).

RealityVoid 10y ago
My experience mirrors yours, but I only worked on this for @ 5 months give or take and didn't really have a big tech team to back me up. The Qualcomm platform, as far as my understanding goes, still uses the PX4Flow. I don't know what packages they have for vision but it uses ROS so I assumed it had some ROS package powering it. Px4Flow on its own is insufficient, but I was under the impression that it is able to calculate velocity even with variable distance camera-to-surface, they do have the ultrasonic sensor there and apparently, adding a lidar greatly improves performance.
- fest 10y ago
  Qualcomm platform has different camera and is directly connected to SoC, so I doubt they share anything with PX4Flow. PX4Flow can calculate velocity when camera-to-surface distance varies, no doubt about it. The problem is when single frame contains features at various distances (e.g. lidar/sonar measures distance to ground, but half of the frame also sees table which is a lot higher).

by fest 10y ago

RealityVoid 10y ago

My experience mirrors yours, but I only worked on this for @ 5 months give or take and didn't really have a big tech team to back me up. The Qualcomm platform, as far as my understanding goes, still uses the PX4Flow. I don't know what packages they have for vision but it uses ROS so I assumed it had some ROS package powering it. Px4Flow on its own is insufficient, but I was under the impression that it is able to calculate velocity even with variable distance camera-to-surface, they do have the ultrasonic sensor there and apparently, adding a lidar greatly improves performance.

fest 10y ago
Qualcomm platform has different camera and is directly connected to SoC, so I doubt they share anything with PX4Flow. PX4Flow can calculate velocity when camera-to-surface distance varies, no doubt about it. The problem is when single frame contains features at various distances (e.g. lidar/sonar measures distance to ground, but half of the frame also sees table which is a lot higher).