Project for COMP9517 ( Computer Vision )

Tasks

Task 1

• The system will report:
  • The unique number of pedestrians detected
  • The path followed by each pedestrian

Task 2

1. Program will ask you to draw the bounding box
2. After drawing the bounding box, press Enter
3. The system will report:
   • Pedestrians who enter the bounding box
   • Pedestrians who move out of the bounding box

Task 3

• The system will report:
  • The number of pedestrians in group.
  • The number of pedestrians not in group.

How to run:

1. Run download pre-trained weights (.pt file) from grdive and put them in weights directory.
2. Place all images in following path -> Group_Component/sequence/*.jpg
3. Download ckpt.t7 from gdrive and place in deep_sort directory
4. Run python TaskSolver.py --task <TASK_NUMBER>
5. Check python TaskSolver.py -h for other custom arguments

Weights

our weights was trainned with COCO_2017 dataset with only person class (64,115 images).

Results

With 300 epochs, we got mAP 0.644

References

• This project uses YOLOv3-SPP model to detect pedestrians, as implemented by ultraltyics team.
• Deepsort implemented by ZQPei to track pedestrians.