Inference server for truck detection (waste reduction project) at Code for Israel
Inference Server
- Build and deploy a Flask app using Docker.
- Development framework: PyCharm
- Three json files:
- cameras.json
- credentials.json
- config.json
Python Classes
- Config - All configuration variables, extracted from the three json files, needed for the inference server (Cameras’ urls, AWS credentials, directories names, etc...).
- VideoCapture
- Initiate access to camera - Start the connection to camera (GET 200 response code instead of 403) by using Selenium.
- Read frames - Reading frames from camera, on a thread, by a given frame rate.
- Scraper - Saving a frame with image name format (camera’s ID, serial number, date and time) into a local input folder.
- Model - Using the model’s weights to infer images. Saving output images and labels (of exist) into a local output folder.
- Files
- A logger instance for alerts logging and other metadata logging.
- Creating input and output folders.
- Deleting local folders and images.
- Uploading and downloading files from a s3 bucket.
Automatic scheduler
- Sends output, in json format, to the backend server every x=0.5 seconds (x can be over 0.2 seconds).
- Uploads every 5 minutes the log file to s3 bucket.
Inference time on local server
- Scraping and saving a frame: 1 - 14 ms
- Inferring and saving outputs (images + labels): 180-190 ms
- Total time: around 200 ms (0.2 seconds).
