A system for using RFID tags and computer vision to track multiple mouse movement within a small environment. For more detailed information, please check the wiki! For a brief overview, read on!
Uses 4 ID-20LA RFID tag readers, which then interface with the Raspberry Pi with serial communication. The tag readers are set up in the four corners of the cage, to minimize electromagnetic interference. This system is constantly running in parallel with the Overhead Camera, recording each tag read with respect to frames of the video.
Records videoa at 15.0 fps, converts to grayscale, and saves for use in the object detection system.
An object detection system trained in-house using https://pjreddie.com/darknet/yolo/. Stands for "You Only Look Once" - this system is very fast and accurate, and only fails when mice climb completely atop one another, or are obscured from the camera's vision. The information from the object detection system feeds into the Mouse Trackers.
When the detection system sees as many objects as there are mice, then the system simply tracks based on Euclidean distance, choosing the closest pairs of detections until all mice and detections are covered. Since the framerate is faster than the mouse movement, this is a reliable method for tracking all mice. When mice are lost, a mouse is removed from the main list and a "dummy mouse" is added in place - so, if two mice are lost, then they are replaced with dummy mice henceforth. Once the mice return, the RFID system is checked continuously. Once we have found a frame in which a mouse has been read by the RFID sensors, the dummy mice are classified and the system is retroactively corrected. In this way, even though we may not be certain of which mouse is which on every frame moving forwards in time, once the system has finished running, all mice movement will have been completely tracked.
RFID validation is used to correct and update the current state of the system. If an identity swap occurs that is not handled by the preventative destruction method, then there must be countermeasures. Thus, if a mouse identified by the system as tracked walks over an RFID sensor and is shown to be a different tag, then its data must be removed up to the last validation point, i.e. the last point it was confirmed by an RFID sensor.
By combining these methods, we can confirm no identity swaps and an 80%+ coverage rating for this system.
- https://github.com/Ebonclaw/Mouse-Wearable-Tech---RFID-and-Localization-Grid-Computer-Vision-Enhancement - original inspiration
- https://github.com/AlexeyAB/darknet - object detection
- https://github.com/jamieboyd/RFIDTagReader - tag reader