My research interest sits where algorithms meet limited compute, unreliable connectivity, uncertain environments and users who cannot tolerate fragile demos.
Accuracy in isolation is not enough. A useful system must also manage latency, power, model size, sensor uncertainty, privacy, affordability and the clarity of its feedback.
Offline computer vision for mobility and environmental awareness, designed around privacy, response time and affordable hardware.
Perception and control for small rovers operating with limited sensing, imperfect terrain and explicit mission rules.
Risk modelling that converts historic crashes, time, weather and road context into interpretable location-hour risk estimates. The goal is decision support, not a false promise of certainty.
Each stage exposes different failure modes. Skipping one usually moves the failure downstream.
Define the user, environment, failure cost and measurable outcome.
Collect data, annotate carefully and inspect where the dataset lies.
Train baselines, test alternatives and measure the real trade-offs.
Profile on target hardware, observe field failures and iterate.
These recorded project results show the progression across model choice, dataset quality and training iteration. They should be read as project-specific results, not universal benchmarks.
| Iteration | Model | Recorded result | Lesson |
|---|---|---|---|
| Baseline | YOLOv8n | 29% | Speed alone could not compensate for weak coverage. |
| Expanded run | YOLOv8n | 63% | Training and data improvements materially changed the outcome. |
| High-capacity run | YOLOv8m | 87% | Capacity helped, but created a harder edge-deployment problem. |