Most public discussions about autonomous vehicles focus on highway driving. This project looked at something messier: low-speed urban transit corridors where buses and trams share space with cyclists, delivery workers, and pedestrians moving unpredictably across unmarked paths.
The problem this project investigates
Existing pedestrian detection benchmarks are heavily weighted toward clear daylight conditions and marked crosswalks. When the same models run against footage from a downtown transit corridor at dusk, during light rain, with partial occlusion from parked vehicles, detection confidence drops in ways that are not obvious from headline accuracy numbers.
The project built a small-scale testing environment using a Raspberry Pi 4 with a camera module, a pre-trained YOLOv8 model, and custom video clips recorded at three locations in a mid-size Canadian city. Detection rates were logged frame-by-frame and compared across lighting and occlusion conditions.
What the data showed
Partial occlusion — a pedestrian visible from the waist up behind a parked car — caused the largest drop in detection confidence, more than low light alone. Combining both conditions produced results that would be concerning in a real deployment scenario, though the sample size here is too small to draw broad conclusions.
- Model tested: YOLOv8n (nano variant, optimized for edge hardware)
- Locations: 3 urban transit stops, recorded across morning, midday, and dusk sessions
- Frames analyzed: approximately 9,400 across 6 recording sessions
- Key finding: occlusion plus low contrast background reduced detection rate more than either factor alone
This is not an argument against autonomous transit. It is an argument for being specific about what current models can and cannot do reliably before deployment decisions are made.
The project includes a written analysis of three published safety frameworks and a proposal for supplementary sensor configurations that could reduce the gap identified in the tests.