Road Traffic Forecast

Road traffic flow detection on video data collected from two sources

• Videos captured from tablets mounted on ODOT trucks ( I can't share the data here as it is not available for public use).
• Videos obtained from the Caltrans Measurement System (CMS) in the state of California 2GB Samples.

Introduction

The purpose of this project is to develop a system that can extract traffic information from live video streams. This project consists of two parts mainly. The first part is extract traffic data from videos. The second part is to train forecasting models using LSTM and GRU Neural Networks.

project Walkthrough

• install the project dependencies inside requirements.txt to your working environment.
• place the videos you want to work on in the Video_Data_Samples
• run VehicleCounterYOLO.ipynb if you do not care about real time data extraction the code is based on YOLO V3 published by Redmon and Farhadi(2018) paper . The counting API takes around 0.9 second on NVIDIA GeForce GTX 1050 to process one video frame. very slow to be aplied on real time sources but much more accurate. (tested with around 96% accuracy). SSD object detector was also tested but it signifacantly struggled with smaller objects (tested with around 92% accuracy). for the SSD the time it required to run to forward a frame was around 0.1 second which a significant improvment compared with YOLO but still very slow for real time object detection specially for low processing power components that are typically used for Road Monitoring systems, for example Beagle Board, Raspberry Pi. Before running the code download the yolo-coco model weights from this Link and put it inside yolo-coco directory.
• run VehicleCounterOpenCV.ipynb for real time data extraction. accuracy around 85% , frame processing time is less than 0.03 seconds.

• The counter codes will automatically extract the data to a csv format file which can be used to train the RNN models by running traficForecast.ipynb. you can skip the training process and load models from models directory.

Results

Metrics	MAE	MSE	RMSE	MAPE	R2	Explained variance score
LSTM	7.097	93.554	9.672	17.059%	0.942	0.942
GRU	7.267	95.928	9.794	19.952%	0.940	0.941

• To test the models with your own input run the webapp by going to the webapp directory and typing flask run from command line. The demo webpage will be running on: http://127.0.0.1:5000/
• Currently unable to deploy a demo on heroku as the dependencies were larger than the maximum size allowed for free access.

Website Demo Images:

Website Demo video: YouTube Link