Hey there!
This is how you can develop a deep learning computer vision system quickly in about two days.

1) Come up with a use case
   * Use AI to monitor detect when cars cross into bike lane
     - To be seen in broader context. Real use case is sth like:
       + monitor a large set of crossings, roads, etc
       + identify most dangerous spots and violations etc
       + act, impose fines, ..

2) Get data
   * google "Watch CCTV camera feed"
     - 2nd hit is http://www.insecam.org/ 
   * Browse a view cameras, find this one:
     http://www.insecam.org/en/view/891271/
   * capture images from CCTV
     - google "stack overflow capture camera feed"
     - 2nd hit: https://stackoverflow.com/questions/55957298/how-to-read-live-video-feed-or-video-on-demand-feed-in-python
     - write: capture-data.py

3) label data
   * use CVAT https://cvat.org

4) Train algo
   * follow Pytorch tutorial https://pytorch.org/tutorials/intermediate/torchvision_tutorial.html#finetuning-from-a-pretrained-model 
     - had shown and posted code on barcode bounding box detection
     - here exchange faster-rcnn (train-faster-rcnn.py) with mask-rcnn.py
     - train-mask-rcnn.py
   * activate your pytorch environemtn (see https://pytorch.org/ for installation)
     env-pytorch\Scripts\activate
   * train with batch size fitting memory of your GPU, adjust learning rate like 1/N_GPU, 1/batch_size
     0.02 / 16 = 0.00125
     python train-mask-rcnn.py -b 1 --lr 0.00125

5) Be happy