![[personal profile]](https://www.dreamwidth.org/img/silk/identity/user.png){kind=small}
![[identity profile]](https://www.dreamwidth.org/img/silk/identity/openid.png){kind=small}
I spent the day working on my sensorimotor project, which is coming along surprisingly well.
This is the eye simulation hardware. It's basically just a camera mounted on a motor with a shaft encoder. It also has a accelerometer, but that isn't being used at the moment. The whole thing sits on a lazy susan for easy rotation. It already had code for simulating saccades (jumps from one position to another) along with a good amount of utility functions, but the only person who knew how to get it running went back to Sweden last December. Part of my project writeup will have to be a proper set of docs for the thing.
For the object tracking part of the project, I needed targets to move around in front of the camera. So I made Charles Babbage and Alan Turing puppets. I certainly don't make them dance up and down or talk in squeaky voices. That would be unprofessional.
The system works using SIFT features, which are basically points in an image that always look the same no matter how the image is oriented, how big it is, or (to a lesser extent) how it was lit. This is the Babbage image with all of its features marked.
Features are then found in frames from the camera. They are matched to features from the training image. Note that most aren't matched, but because there were ~750 to work from, we still get plenty of hits. The code then uses these to determine the center of the object and moves the camera in that direction.
Next steps: The camera motion is currently quite naive and unstable if not heavily dampened. I need to implement something like a proper PID control loop for that, but I'm going to talk to the prof first to find out what is the most biologically correct model to use. I also want to try adding an interface to dynamically choose a region to track instead of using static input images. But really, it's all coming together quite well. With another month left, I can safely move this out of the 'omg stress' category.
This is the eye simulation hardware. It's basically just a camera mounted on a motor with a shaft encoder. It also has a accelerometer, but that isn't being used at the moment. The whole thing sits on a lazy susan for easy rotation. It already had code for simulating saccades (jumps from one position to another) along with a good amount of utility functions, but the only person who knew how to get it running went back to Sweden last December. Part of my project writeup will have to be a proper set of docs for the thing.
For the object tracking part of the project, I needed targets to move around in front of the camera. So I made Charles Babbage and Alan Turing puppets. I certainly don't make them dance up and down or talk in squeaky voices. That would be unprofessional.
The system works using SIFT features, which are basically points in an image that always look the same no matter how the image is oriented, how big it is, or (to a lesser extent) how it was lit. This is the Babbage image with all of its features marked.
Features are then found in frames from the camera. They are matched to features from the training image. Note that most aren't matched, but because there were ~750 to work from, we still get plenty of hits. The code then uses these to determine the center of the object and moves the camera in that direction.
Next steps: The camera motion is currently quite naive and unstable if not heavily dampened. I need to implement something like a proper PID control loop for that, but I'm going to talk to the prof first to find out what is the most biologically correct model to use. I also want to try adding an interface to dynamically choose a region to track instead of using static input images. But really, it's all coming together quite well. With another month left, I can safely move this out of the 'omg stress' category.
no subject
(no subject)
no subject