Pan/tilt mount for a cell phone

Sat, 23 May 2020 00:00:00 +0200

I think it’s important to also post about the projects that didn’t turn out. Pudel was a project to automatically track and video a model plane using a cell phone camera.

The pan and tilt components were modeled in OpenSCAD and 3D printed. The pan unit had a standard 1/4-20 tripod mount and the tilt unit had a Samsung A40 mount that I’ve used on a few other projects.

Both were driven by standard HS-422 servos with a Trinket M0 for I/O and a 2S LiPo through a 6V BEC for power.

The I/O board ran CircuitPython with custom firmware that takes MIDI-like messages as servo commands. It implemented an acceleration limiter and velocity limiter to limit the jerk, as the servos are capable of 60 deg/0.2 s which is much faster than the plane would move.

The host side was going to be similar to Kieferbot and use ROS with an IO specific node, OpenCV based detection, and a PID controller to keep the plane centered in the frame.

So why did it fail?

The servos aren’t intended for precise, slow movement and hence had mild jerk.
There was too much slack in the mechanics, and the phone was mildly heavy, so mild jerk caused the phone to move about too much.
The plane is too small to resolve on a wide angle lens, so tracking was unlikely to work.

If I wanted to take it further, I can imagine some improvements:

Take the load off the pan servo. The tilt unit sits directly on top of the pan servo shaft and there’s not much material to sit on which increased the shake. Perhaps add a turntable style plates that most of the downwards force could go through.
Decrease the arm length out to the phone. It’s 30 mm and reducing to 20 mm would reduce the load on the tilt servo.
Try a phone with optical stabilization.

Tracking using a camera

Sun, 19 Apr 2020 00:00:00 +0200

Here’s Kieferbot being tracked through a camera:

It took a few iterations but I settled on matching the colour of the floor to create a floor mask, then findCountours to find the contours as a tree, and then searching the children of the root contours. As the bot is black, the children were matched against a ‘black pixels’ mask, and the contour with the most black pixels picked.

The camera is an Android phone running IP Webcam and sharing the back camera as MJPEG over HTTP over Wifi at 10 FPS. It could be higher, but the laptop doing the image processing has trouble keeping up with 30 FPS. The latency looks good - visually less than 200 ms.

The tracking seems reliable. If the bot goes off the bottom of the frame then other black areas are picked, but this can be detected by checking for a jump in position.

Next step is to integrate the camera position, phone sensors, and wheel velocity to get a controllable position.

ros | Blue Duck Valley Rd

Pan/tilt mount for a cell phone

Tracking using a camera