o-gent / uas-nyx Goto Github PK

state time recording behaviour isn't correct, fix it while/and add tests

• Move parameter setting loop from main.py to a function within mission.py
• Have a reset parameters method

Method within Mission, mission.py

write this and test accuracy and performance changes

Adjust image recognition parameters, methods to increase % pass rate on image rec tests

Thresholding & edge recognition is the main area of failure

Ideas

• tune thresholding parameters
• change thresholding method
• edge recognition parameters

Make a new branch to propose parameter changes,
create a merge request with test results

given a multiple waypoint locations, colours and characters; assess which are the same waypoint and for individual waypoints find a most likely value for location, colour and character.

Using state.py

• Implement a python parser for .params file
• Ensure red-list params aren't loaded

Formalise split between segmentation, OCR and other data processing

just need to get them from the config dictionary

Use waypoints from comp day

https://docs.opencv.org/3.4/d3/dc1/tutorial_basic_linear_transform.html
https://stackoverflow.com/questions/57030125/automatically-adjusting-brightness-of-image-with-opencv
https://stackoverflow.com/questions/32609098/how-to-fast-change-image-brightness-with-python-opencv

a modification of this should do it:
https://stackoverflow.com/questions/50066619/python-non-blocking-write-csv-file/50072105

camera takes image
-> run image recognition on RAM image
-> write to file in background (should not affect main thread speed) for dataset collection and checking performance

• loading real parameters will destroy the SITL
• Some params on a live aircraft can't be set, use a redlist

Instead of using a bash console, use much more efficient hex codes for coms ?

Need a % pass on the dataset to compare image recognition code changes

Rather than transferring them using a USB

Produce a more representative dataset

Ideas:

• blend target with surroundings?
• remove images which don't represent our situation (buildings present etc)
• new representative images
• scale targets correctly for the height we are flying at

Setting SITL params to real vehicle will break it

Aircraft will make a best effort to reach drop point but may miss, so drop bomb when perpendicular to point

see bomb_computer.py
Currently runs a relatively simple model to predict drop location