Description

Emergency Enabling of Highway Emergency Lanes Based on High Precision Simulation and Deep Learning. This project is for the 2024 Graduate Student Mathematical Modeling Competition E Question. This project is inspired by and references the work done in win10_yolov5_deepsort_counting.

The simple schema

There are four video observation points on a road section (length about 5000m, travel lane 2 + emergency lane 1)
Explore the doc »

View Demo · Report Bug · Request Feature

Solving Problem: How to Construct a Highway Traffic Congestion Model There are four video observation points (see Schematic 1) on a certain road section (length about 5000m, travel lane 2 + emergency lane 1): (1) For the data provided in the question, count the change rule of traffic flow parameters over time at the four observation points (finer statistics can be more conducive to future modeling); (2) the establishment of traffic flow congestion model, using the traffic flow in the four observation points of the basic parameters (traffic density, flow, speed, etc.) and the road situation (two lanes), give from the third point to the fourth blockage before the point between the road section may (no congestion) appear sustained (for example, the duration of half an hour) the state of the real-time warning of the congestion (for example, congestion 10 minutes before the warning) and the basis;

Content

• Built With
• About The Project
  • Dependencies
  • Installation
• Project Tree
• Usage
• Roadmap
• Author
• Acknowledgments

Built With

• Python
• scikit-learn
• AutoGluon

Getting Start

Dependencies

• autogluon = "1.1.1"
• matplotlib = ">=3.2.2"
• numpy = ">=1.18.5"
• opencv-python = "4.9.0.80"
• Pillow = ">=7.1.2"
• PyYAML = ">=5.3.1"
• requests = ">=2.23.0"
• scipy = ">=1.4.1"
• torch = ">=1.7.0"
• torchvision = ">=0.8.1"
• tqdm = ">=4.41.0"
• tensorboard = ">=2.4.1"
• pandas = ">=1.1.4"
• seaborn = ">=0.11.0"

Installation

1. install poetry
2. Clone the repo
3. install python and make sure that the version >=3.10
4. install autogluon

conda install -c conda-forge mamba
mamba install -c conda-forge autogluon "pytorch=*=cuda*"
mamba install -c conda-forge "ray-tune >=2.6.3,<2.7" "ray-default >=2.6.3,<2.7"  # install ray for faster training

5. install other dependencies:

pip install -r requirements.txt

6. Install the project with poetry

poetry install

Project

Tree eg:

traffic-congestion-detection
├── notebook
├── figure
│   ├── raw
├── res
│   ├── configs
├── src
│   ├── models
│   ├── results
│   ├── traffic
│   │   ├── chart
│   │   ├── detect
│   │   ├── predict
│   │   ├── analysis
├── tests
├── .gitignore
├── README.md
├── pyproject.toml
├── requirements.txt
└── LICENSE.txt

Usage

Before running the project, ensure that the current working directory is set to the project directory. And then you can get the video detection result by the following command:

poetry run detect

Next you can draw the result by

poetry run draw

Or you can re-calculate the congestion information by:

poetry run calculate

Lastly, you can view the prediction results from time series models by:

poetry run predict

Roadmap

Step1: YOLOV5+DeepSort recognizes vehicles, refer to win10_yolov5_deepsort_counting, in which we modified its recognition range and added separate recognition statistics for different lanes (including emergency lanes).

Subplot 1: 107 port	Subplot 2: 105 port
Subplot 3: 108 port	Subplot 4: 103 port

Step2: Calculate the traffic information, you can see all the features in traffic_info.md and all the results in result-folder line_page.png abd hist_page.png. The following is some main results:

Number of all vehicles at all times	Densities of all vehicles
Densities of all vehicles in the emergency lane	Speed of all vehicles

Step3: Predict the congestion with 2 mode: Self-mode and Global-mode.

The self-mode means that the detection point predicts the traffic congestion in the next ten minutes based only on the road traffic data of the past period of time at that point (ignoring the data of other detection points). In this model: to predict the congestion at site 103 for the next 10 minutes, the time series model only accepts data from the 103 detection points, where the degree of traffic congestion is used as the target variable, and the covariates are the road condition data (including the flow, density, and speed of the main lanes, the overtaking lanes, and the emergency lanes, as well as the number of cars, trucks, buses, etc., on the road).

The global-mode means that the simultaneous reference to the traffic road data of the four detection sites in the past period of time to predict the traffic congestion in the next ten minutes. For example, to predict the next 10 minutes of congestion at station 103 target variable, the time series model simultaneously accepts the road condition data of the four monitoring points as covariates, and at the same time introduces the poi values of the four stations as static variables for the training of the time series model.

Self mode

Global mode

Result

Self mode

Global mode

Mode	Model	MASE	RMSE	MAPE	SMAPE	MAE	RMSSE	WAPE	WQL	SQL
Self	SeasonalNaive	0.056	0.066	0.398	0.315	0.056	0.066	0.338	0.301	0.050
	RecursiveTabular	0.076	0.091	0.538	0.660	0.076	0.091	0.456	0.347	0.057
	DirectTabular	0.051	0.059	0.476	0.332	0.051	0.059	0.310	0.258	0.043
	CrostonSBA	0.040	0.048	0.366	0.273	0.040	0.048	0.242	0.181	0.030
	NPTS	0.059	0.073	0.598	0.360	0.059	0.073	0.357	0.343	0.057
	DynamicOptimizedTheta	0.045	0.055	0.350	0.273	0.045	0.055	0.269	0.212	0.035
	AutoETS	0.047	0.057	0.365	0.285	0.047	0.057	0.283	0.225	0.037
	AutoARIMA	0.045	0.055	0.391	0.289	0.045	0.055	0.273	0.240	0.040
	Chronos[base]	0.047	0.057	0.431	0.308	0.047	0.057	0.286	0.240	0.040
	TemporalFusionTransformer	0.056	0.065	0.481	0.343	0.056	0.065	0.338	0.281	0.046
	DeepAR	0.064	0.074	0.510	0.367	0.064	0.074	0.386	0.323	0.053
	PatchTST	0.035	0.044	0.341	0.248	0.035	0.044	0.209	0.172	0.029
	WeightedEnsemble	0.037	0.046	0.341	0.248	0.037	0.046	0.223	0.189	0.031
Global	SeasonalNaive	0.010	0.010	0.182	0.167	0.010	0.010	0.182	0.645	0.037
	RecursiveTabular	0.171	0.171	2.994	1.199	0.171	0.171	2.994	2.045	0.117
	DirectTabular	0.119	0.119	2.087	1.021	0.119	0.119	2.087	1.819	0.104
	CrostonSBA	0.033	0.033	0.576	0.447	0.033	0.033	0.576	0.402	0.023
	NPTS	0.143	0.143	2.500	1.111	0.143	0.143	2.500	2.345	0.134
	DynamicOptimizedTheta	0.002	0.002	0.037	0.036	0.002	0.002	0.037	0.296	0.017
	AutoETS	0.010	0.010	0.183	0.168	0.010	0.010	0.183	0.147	0.008
	AutoARIMA	0.011	0.011	0.186	0.170	0.011	0.011	0.186	0.276	0.016
	Chronos[base]	0.127	0.127	2.228	1.054	0.127	0.127	2.228	2.098	0.120
	TemporalFusionTransformer	0.157	0.157	2.747	1.157	0.157	0.157	2.747	2.590	0.148
	DeepAR	0.183	0.183	3.202	1.231	0.183	0.183	3.202	3.111	0.178
	PatchTST	0.129	0.129	2.263	1.062	0.129	0.129	2.263	2.202	0.126
	WeightedEnsemble	0.172	0.172	3.007	1.201	0.172	0.172	3.007	2.810	0.161

Author

Haowei Wang

AlexLiu

Fan Wang

License

This project is licensed under the MIT License. For more details, please refer to LICENSE.txt

Acknowledgments

• GitHub Pages
• isort
• ruff