An API for IP camera video streaming and object detection using DETR(DEtection TRansformer) or Faster-rcnn. Both of them uses resnet50 as backbone. DETR is about 4x faster but Faster-RCNN is more precise on real scenes. Built with FastAPI and PyTorch.
uvicorn app:app
full demo: https://youtu.be/fFWtUqfK36M
This API was built using Intelbras cameras, this variables are related to the rtsp url, if you use a different type of camera, there's a variable CUSTOM_URL where you can just pass the entire url for your cameras at rtsp.
If you use Intelbras Cameras, you can just use the following Create a .env file with the following variables:
USER=
PASSWORD=
CAMERA_IP=
RTSP_PORT=
PORT=
ENVIRONMENT_NAME=
MODEL=
GPU=- USER: Your user within your camera system.
- PASSWORD: Your password within the system.
- CAMERA_IP: Your camera IP.
- RTSP_PORT: 554 as default, but you can change if you changed in your system.
- PORT: Api port, 8000 as default.
- ENVIRONMENT_NAME: Will disable swagger and redocs for FastAPI if starts with "production". Default is None.
- GPU: Default is false, 1 to use cuda.
- MODEL: default is DETR, FASTER-RCNN available.
- /cam
Params: channel
Description: open the camera x (1,2,3,4...).
Example: 127.0.0.1:8000/cam?channel=5
- Demo DETR:
- learned positional encoding (instead of sine)
- positional encoding is passed at input (instead of attention)
- fc bbox predictor (instead of MLP)
- A faster-rcnn with a resnet50 backbone as the cnn, available at torchvision and trained using COCO dataset.
- Custom set of categories on detection.
- Track desired objects in a specific space of time.
- Interface (now its just buffering the camera frames into the model and displaying at the cam route).
- Implement best DETR and see if its still worse than faster-rcnn on accuracy.