This project focuses on real-time vehicle detection for self-driving applications using the YOLOv5 object detection framework.
The model identifies five vehicle classes — Car, Truck, Bus, Motorcycle, and Ambulance — using the Vehicles OpenImages dataset.
Developed as part of the Master’s in Management Information Systems / Machine Learning program at the University of Arizona.
The goal of this project is to train and compare different variants of the YOLOv5 model — Nano, Small, Medium, and Medium (Frozen Layers) — to analyze their trade-offs between speed, accuracy, and computational efficiency.
The models were fine-tuned on a custom dataset, evaluated on unseen data, and tested on real-world images and videos for autonomous vehicle perception.
| Model | Parameters (M) | Precision | Recall | mAP@0.5 | mAP@0.5–0.95 | Training Time (hrs) |
|---|---|---|---|---|---|---|
| YOLOv5n (Nano) | 1.77 | 0.441 | 0.466 | 0.441 | 0.295 | ~1.0 |
| YOLOv5s (Small) | 7.02 | 0.489 | 0.615 | 0.540 | 0.364 | ~1.9 |
| YOLOv5m (Medium) | 20.9 | 0.606 | 0.647 | 0.623 | 0.467 | ~0.12 |
| YOLOv5m (Frozen 15 Layers) | 20.9 | 0.645 | 0.606 | 0.648 | 0.466 | ~0.10 |
Best Performing Model: YOLOv5m (Frozen 15 Layers) — achieved the best balance between accuracy, speed, and stability.
YOLOv5 models were validated on unseen data to measure performance across classes:
| Validation | Inference |
|---|---|
 |
- Used Vehicles OpenImages dataset (627 images, 1194 annotations)
- Split into train, valid, and test sets
- Converted labels to YOLO format using a YAML configuration file
Trained multiple YOLOv5 models using Google Colab:
- YOLOv5n.pt → Nano (lightweight baseline)
- YOLOv5s.pt → Small (balanced model)
- YOLOv5m.pt → Medium (best accuracy)
- YOLOv5m (Frozen) → Transfer learning, 15 layers frozen
Each model was trained for 15 epochs with TensorBoard monitoring.
- Plotted metrics: precision-recall, F1, and confusion matrices
- Visualized predicted bounding boxes on validation data
- Performed inference on unseen images and videos
- Saved detection results to the
runs/detect/directory
- Increasing model size improved detection accuracy and recall.
- The YOLOv5m (Frozen Layers) model achieved the best trade-off between training time and precision.
- Transfer learning and selective freezing improved convergence speed and model stability.
The YOLOv5m (Frozen) model is ideal for real-world self-driving systems, balancing efficiency and reliability for real-time object detection.
- Apply pruning or quantization for edge deployment.
- Perform advanced hyperparameter tuning to boost accuracy.
- Add new classes (e.g., pedestrians, bicycles).
- Include nighttime and adverse weather conditions.
- Add GPU acceleration and object tracking (e.g., DeepSORT) for real-time video analysis.
- Extend YOLOv5 to new domains (e.g., agriculture or healthcare).
- Apply Explainable AI (XAI) methods such as Grad-CAM or SHAP.
| Category | Tools |
|---|---|
| Language | Python 3.10 |
| Framework | PyTorch, Ultralytics YOLOv5 |
| IDE / Environment | Google Colab |
| Visualization | Matplotlib, TensorBoard |
| Dataset | Vehicles OpenImages (via Roboflow) |
| Version Control | Git, GitHub |
self-driving-yolov5-vehicles/
│
├── .gitignore
├── README.md
├── requirements.txt
│
├── data/
│ ├── Vehicles-OpenImages.zip
│ └── inference_data_yolov5.zip
│
├── notebooks/
│ ├── Self_Driving_Car_Object_Detection_Using_YOLOv5.ipynb
│ └── self_driving_car_object_detection_using_yolov5.py
│
└── runs/ (optional)
Ajibola Jeremiah Dedenuola
Master’s in MIS/ML, University of Arizona
- Ultralytics for the YOLOv5 framework
- OpenImages Dataset for high-quality vehicle images
- Google Colab for GPU/TPU compute resources
This project is distributed under the MIT License.