| Departments |
Computer Science
|
|---|
| Description | In this MSc project, students will develop advanced machine learning techniques for detecting potholes in dashcam-style videos recorded from vehicles around Swansea and UK roads. The core challenge is achieving reliable detection across diverse lighting conditions (daylight, dusk, night, rain, glare), with proper bounding box annotations for precise localisation. A new dataset will be created from real-world dashcam footage, featuring potholes at vehicle height with GPS metadata for validation.
This addresses the critical need for proactive road monitoring: UK potholes cause over £1.7 billion in annual vehicle repairs and contribute to a £17 billion road backlog, exacerbated by councils' slow, complaint-only responses. If successful, the model could power a simple user app where drivers flag detected potholes to local authorities with one-tap GPS sharing.
Students will review state-of-the-art video object detection methods, implement and compare models (e.g., YOLO variants, temporal CNNs), annotate the custom Swansea dataset meticulously, and evaluate quantitative metrics (mAP, precision/recall) alongside qualitative analysis of failure cases. Driving students are encouraged to contribute video data. Contact Dr. Gary K. L. Tam for details. |
|---|
| Preparation | Interesting students should be aware of the challenge and difficulty (e.g., long training time, understanding complex pytorch codes) of deep learning.
You will need a better graphics card (e.g., RTX3090, 4090, 5090) for deep training. Google colab may be cheaper however but observe the limitations.
Suits all students apart from “non-technical” (i.e. good for MSci, MEng, BSc, Software Engineering, MSc CS, MSc Adv CS, MSc SoftTech etc). |
|---|
| Project Categories |
Artificial Intelligence (AI), Data Science |
|---|
| Project Keywords |
Computer Vision, Machine Learning |
|---|
Level of Studies
|
| Level 6 (Undergraduate Year 3) |
yes
|
|---|
| Level 7 (Masters) |
yes
|
|---|
| Level 8 (PhD) |
yes
|
|---|