hju32 / test_module

Target detection code for Gatech DBF 2018 Medical Express Challenge

This project aims to be a robust algorithm to detect targets

Author

Author: Eric Fang (CMU), Nick Zhang [email protected]

TO-DOs - Tickets

Description	Status	comment
Reduce box drawing time	open
Go-pro undistortion	Suspended	we will do that for real computer
fix white balance (find ways how)	closed	update: our GoPro can't do that

Pipeline:(single image, no access to historic data)

Pre-adjustment: (exp) sharpening? Undistort

Tracking: Mask out prevalent color pixels(histogram, greenest) Mask small region of interest (area with unique colors) Local normalization/re-whitebalance in each ROI (prevent bias,lighting, WB) Edge detection, find lines, find crossing -> output.coordinate One crossing each ROI

Then, score each crossing, based on following criterion: Check length of crossed edges, does it look like our chessboard? Check color near crossing, are they alternating black/white?

Issues

When close to target, whites in target will be considered mode, therefore ROI may filter target out. This appears in clip1( 1:09-1:20 )

Some points get picked up in the pre-processing routine as flashing dots that appear intermittently from frame to frame This appears in clip1( 20-30, 49-52 )

Suggestions

Change target to non-directional pattern, like concentric circles

Notes

Gopro 1080p delivers usable source at 100ft altitute.