Motion detection with 100% recall on Jetson Nano

KamiCreed commented 9 months ago

Motion detection should be the first thing to process to prevent false negatives from the automated counting later in the pipeline, allowing manual review and potential incorporation into future datasets. This means 100% recall with low enough false positives. Keep in mind to build all of this for the Jetson Nano (ARM64 system), and the underwater cameras send out RTSP input. Check the note in utils/jetson/README.md for help using a previously setup Jetson.

[x] Save continuous video locally in the harddrive in 30 minute chunks just in case
[x] #75
[x] Determine motion detection algorithm
[x] Background subtraction algorithm (MOG2, CNT, etc.)
[x] Erosion and Dilation
[x] #85
[ ] Test on annotated dataset and figure out how to get the recall metric
[ ] #67
[ ] Setup alerts with docker-compose when it goes down maybe with healthchecks.io?

KamiCreed commented 9 months ago

Motion detection script is basically completed but still need to troubleshoot why the video clips come out empty when run on the Jetson Nano.

KamiCreed commented 9 months ago

Jetson Nano must use the gstreamer pipeline when writing videos. Additionally, it must use nvvidconv and nvv4l2h264enc to use the GPU and be fast enough in encoding the frames as shown in utils/jetson/motion_detect_stream.py.

Salmon-Computer-Vision / salmon-computer-vision

Motion detection with 100% recall on Jetson Nano #66