Towards End-to-End Traffic Congestion Estimation Using Learned ROI and Vehicle Object Dynamics

Implementation of the ISDS2025 paper "Towards End-to-End Traffic Congestion Estimation Using Learned ROI and Vehicle Object Dynamics". The codebase reproduces the full pipeline described in the paper: lightweight road-marking segmentation, geometry-aware ROI derivation, YOLO11s-based vehicle analytics, and Level-of-Service (LoS) congestion classification tailored to motorbike-dominant cities such as Ho Chi Minh City.

Highlights

• Lightweight segmentation DeepLabV3+ head with a ResNet34 backbone enhanced by squeeze-and-excitation and depthwise-separable convolutions, delivering high Dice/mIoU at edge-device speed.
• Automatic ROI discovery Geometry-aware post-processing that converts dashed lane markings into calibrated ROI lines and meter-per-pixel ratios, removing manual camera setup.
• Multi-object analytics YOLO11s fine-tuning, BoT-SORT tracking, and displacement-based speed estimation provide per-lane traffic counts and velocities.
• LoS classification Implements the Highway Capacity Manual thresholds with dual speed tables (two-wheel vs. four-wheel) to output congestion levels AF.
• Production-ready project Structured Python package with configs, scripts, tests, and utilities for training, inference, and evaluation.

Architecture Overview

Figure 1. End-to-End Traffic Congestion Estimation Pipeline.

📁 Repository Structure

Traffic-Congestion-Estimation-ISDS2025/
├── src/                      # Core Python source package
│   ├── config/               # YAML configuration helpers
│   ├── data/                 # Dataset loaders, preprocessors, and utilities
│   ├── models/               # Deep learning model definitions
│   │   ├── segmentation/     # SEResNet34 + DeepLabV3 implementation
│   │   └── detection/        # YOLOv11s wrapper + BoT-SORT for real-time tracking
│   ├── pipeline/             # ROI extraction, speed estimation, and congestion logic
│   └── utils/                # Logging, visualization, and video helper functions
│
├── configs/                  # YAML presets for segmentation, detection, and pipeline
├── scripts/                  # CLI tools for training, evaluation, and inference
├── tests/                    # Unit tests for core components
├── data/README.md            # Notes on dataset structure and RLMD metadata
│
├── README.md                 # Project overview and usage instructions
├── requirements.txt          # Python dependencies

Quick Start

python -m venv .venv
.venv\Scripts\activate  # Windows
pip install -r requirements.txt

Training

Segmentation (RLMD-ready):

python -m scripts.train_segmentation --config configs/segmentation.yaml

• Defaults to the RLMD dataset layout (data/segmentation/images/{train,val}, data/segmentation/labels/{train,val}) with class colors specified in data/segmentation/rlmd.csv.
• Config defaults to the eight RLMD classes highlighted in Table 2 (IDs 4, 5, 6, 7, 8, 9, 10, 16) so only lane-separator markings populate the binary mask; adjust the list if your deployment needs other road symbols.
• Toggle wandb.enabled in configs/segmentation.yaml to stream metrics to Weights & Biases; adjust project/run metadata in the same block.

Vehicle Detection & Tracking:

python -m scripts.finetune_yolo --config configs/detection.yaml

Both scripts rely on the dataset layouts described in data/README.md and export checkpoints to outputs/ by default.

Inference Pipeline

Once checkpoints are ready, update configs/pipeline.yaml with their paths and run:

python -m scripts.run_pipeline --config configs/pipeline.yaml --video path/to/clip.mp4

The script prints per-window LoS summaries and (optionally) saves JSON reports.

Datasets

• Road Marking Segmentation (RLMD example): Drop the RLMD dataset under data/segmentation/ so that images live in images/{train,val} and label PNGs in labels/{train,val}. The provided rlmd.csv palette maps RGB colors to class IDs.
• Detection & Speed: Use a YOLO-compatible dataset with class labels for motorbikes, cars, trucks, etc. Optional speeds.csv can store ground-truth flow speed for validation.

See data/README.md for exact folder expectations, palette handling, and how to override them via YAML.

Results

Component	Metric	Target (Paper)
Road segmentation	Dice / mIoU	75.4% / 73.9%
Vehicle detection	mAP@50	83.1% (YOLO11s finetuned)
Speed estimation	% Error	7.82%
Congestion mapping	Macro F1	85.3%

Actual performance depends on dataset quality, training length, and hardware.

Analysis & Future Work

• Camera shifts: Current ROI lines are computed once per run; integrating temporal ROI alignment (e.g., Temporal ROI Align) would improve robustness to jitter.
• Weather robustness: Augmentations partially mimic rain/fog, but real-world dynamics may require collecting adverse-weather footage.
• Dynamic windows: The fixed 30?s window can be adaptive to capture abrupt congestion changes.
• Visualization: Overlaying ROIs and track info on video output can assist operators; extend run_pipeline.py to use VideoWriter for this.

Citation

If you use this code or the accompanying dataset, please cite the original paper and acknowledge this implementation.

@inproceedings{le2025congestion,
  title={Towards End-to-End Traffic Congestion Estimation Using Learned ROI and Vehicle Object Dynamics},
  author={Le, Nguyen Khang and Tran, Anh Duc and Tran, Thi Minh Tam and Do, Khanh Ngoc},
  booktitle={International Conference on Intelligent Systems and Data Science (ISDS)},
  year={2025}
}

License

MIT License. Dataset licenses may impose additional restrictionsconsult the original sources before redistribution.