Skip to content

Latest commit

 

History

History
51 lines (40 loc) · 1.77 KB

ReadMe.md

File metadata and controls

51 lines (40 loc) · 1.77 KB

Student-teacher Learning for Efficient TrafficCounting

This repository provides the implementation of our paper Automated training of location-specific edge models for traffic counting. The goal of this paper is to count multiple traffic modalities (car, cyclist, pedestrians, and others) with a model that is as small as possible while maintaining a high accuracy. We experimentally show that we achieve similar results as the SToA counting methods with 5x fewer parameters.

Installation and preparation

  1. Clone this repo and prepare the environment
git clone https://github.com/lyn1874/efficient_traffic_count_on_edge_devices.git
cd efficient_traffic_count_on_edge_devices
./requirement.sh

Traffic counting on a toy dataset

python3 inference.py --compound_coef 0 --skip 4

Credits:

TODO

  • traffic counting tutorial
  • update clean code
  • convert model to onnx
  • convert model to coreml
  • debug the coreml model
  • deploy the algorithm on Jetson, report the inference speed
  • simulate the online streaming input

Citation

If you use our code, please cite

@article{LEROUX2022107763,
title = {Automated training of location-specific edge models for traffic counting},
journal = {Computers & Electrical Engineering},
volume = {99},
pages = {107763},
year = {2022},
issn = {0045-7906},
doi = {https://doi.org/10.1016/j.compeleceng.2022.107763},
url = {https://www.sciencedirect.com/science/article/pii/S0045790622000672},
author = {Sam Leroux and Bo Li and Pieter Simoens},
}