Enhanced Safety of Autonomous Driving in Real-World Adverse Weather Conditions via Deep Learning-Based Object Detection

Overview

This repository is based on Claire Zhang’s Master’s Thesis at the University of Ottawa, titled:

"Enhanced Safety of Autonomous Driving in Real-World Adverse Weather Conditions via Deep Learning-Based Object Detection."

This research focuses on one-stage object detection and enhancing YOLOv5 for improved performance in adverse weather conditions. The objective is to boost detection accuracy, robustness, and efficiency to ensure safe autonomous driving in challenging environments.

🚀 Key Contributions

• Enhanced One-Stage Object Detection Framework: Integration of YOLOv5 with multiple lightweight backbones such as ShuffleNetV2, GhostNet, MobileNetv3-Small, and VoVNet.
• Advanced Attention Mechanisms: Incorporation of SE Block, CBAM Block, and ECA Block to improve feature selection and model interpretability.
• Optimized Object Localization: Introduction of Transformer Prediction Heads (TPH) to enhance detection of small and occluded objects in complex scenes.
• Pruning & Quantization Techniques: Model compression strategies for faster inference while maintaining high accuracy.
• Superior Performance in Weather-Adverse Scenarios: Achieved 99.1% mAP (IoU 0.5), significantly surpassing baseline models.

---

📌 One-Stage Object Detection Enhancements

1️⃣ Multi-Backbone YOLOv5 Performance

Model	mAP@50	Parameters (M)	GFLOPs
YOLOv5n	26.2	1.78	4.2
YOLOv5s	34.0	7.05	15.9
YOLOv5x	40.8	86.28	204.4
YOLOv5x-TPH (ours)	63.0	112.97	270.8

2️⃣ Attention-Based Enhancements

Mechanism	Purpose
SE Block	Enhances channel-wise feature selection
CBAM Block	Incorporates spatial and channel attention for improved detection
ECA Block	Refines convolutional feature aggregation

3️⃣ Pruning & Quantization Results

Model	mAP@50	Parameters (M)	GFLOPs
YOLOv5s	34.0	7.05	15.9
[email protected]	27.9	4.59	9.6

---

📂 Installation

pip install -r requirements.txt

🏋️‍♂️ Training & Evaluation

Training

python train.py --data dataset.yaml --weights yolov5s.pt --cfg models/yolov5s.yaml --epochs 300 --batch-size 8 --img 640 --device 0,1

Evaluation

python val.py --data dataset.yaml --weights best.pt --img 640 --device 0

---

📌 Research Background

This research aims to bridge the gap between real-world adverse weather conditions and deep learning-based object detection for autonomous vehicles.

🔬 Research Highlights:

• Investigation of one-stage object detection techniques.
• Evaluation of various lightweight backbones for efficient computation.
• Improved detection in adverse weather using attention-based feature extraction.
• Application of pruning and quantization techniques for real-time deployment.

---

📄 Citation

If you find this work useful, please cite:

@article{zhang2024objectdetection,
  author    = {Biwei Zhang},
  title     = {Enhanced Safety of Autonomous Driving in Real-World Adverse Weather Conditions via Deep Learning-Based Object Detection},
  journal   = {University of Ottawa Thesis},
  year      = {2024}
}

---

🔗 Connect

• GitHub: @catt07986
• ORCID: 0009-0009-7814-9861