Driverless Perception
Efficient vision for Carnegie Mellon Racing
Work on model improvement and knowledge distillation for a YOLO-based traffic-cone detector, balancing useful detection performance with inference constraints.
The engineering problem
A driverless race car must detect colored traffic cones under changing conditions while respecting computational limits. This work focuses on distilling a larger detector into a smaller model, with attention to the tradeoff between accuracy and inference cost.
Current contribution
I’m working with a teammate on the model-improvement and distillation effort. The current direction considers robustness to motion blur and low light, along with efficient inference for embedded deployment.
Evidence to build
The work calls for comparable accuracy and latency measurements across teacher and student models under relevant conditions. Benchmark results, deployment milestones, model specifics, and team materials will be added when verified and cleared for public sharing.