non-overlapping_multiple-camera_tracking

This work references An equalized global graph model-based approach for multi-camera object tracking.

Dataset

This work uses NLPR_MCT dataset
After downloading the dataset, run dat2csv.py to change dat files to csv files.

python dat2csv.py --source PATH/TO/ANNOTATION --dataset DATASET_NO

Method

The association algorithm is iterative min-cost-flow.
The overview approach is shown in below:

Experiments

The evaluation metrics including MCTA, mmes, and mmec.

Dataset 1

Method	MCTA ↑	mmec (inter) ↓	mmes (intra) ↓
EGM	85.25 %	49	66
IMCF (this work)	84.01 %	53	59

Dataset 2

Method	MCTA ↑	mmec (inter) ↓	mmes (intra) ↓
EGM	73.7 %	93	107
IMCF (this work)	85.76 %	60	110

Dataset 3

Method	MCTA ↑	mmec (inter) ↓	mmes (intra) ↓
EGM	47.24 %	80	51
IMCF (this work)	31.77 %	133	82

Dataset 4

Method	MCTA ↑	mmec (inter) ↓	mmes (intra) ↓
EGM	37.78 %	159	128
IMCF (this work)	28.52 %	177	166

How to run?

requirements

• opencv-python==4.5.5
• torchreid

showing videos

python display.py --dataset NO. --data_path PATH/TO/NPLR/DATA --annotation PATH/TO/ANNOTATION --bbox

tracking

run the modules seperately (take sub-dataset 1 as example)

python run_sct.py --dataset 1 --cid 1
python run_sct.py --dataset 1 --cid 2
python run_sct.py --dataset 1 --cid 3
python run_mct.py --dataset 1

NOTE the reid model is downloaded from torchreid model zoo