The perception layer that tells an autonomous go-kart where it can drive

Led perception for a student-built autonomous kart — from road segmentation to a drivable corridor the planner can actually consume, live at 3 Hz.

Role

Perception Lead — owned the full perception stack on a student team

Goal

Camera-only drivable-corridor detection, robust to lighting

Key outcome

94% mAP50 road segmentation, corridor published at 3 Hz

Context

Intelligent Systems Club, University of Michigan-Dearborn

94%

mAP50 (91.4% mAP50-95) — YOLO11 fine-tuned for road segmentation

3×

dataset expansion via augmentation for lighting & viewpoint robustness

3 Hz

drivable corridor published to the planner as live ROS2 topics

$ cat problem.md

A kart needs geometry, not pixels

Raw segmentation masks don't steer a vehicle. The planner needs a drivable corridor — clean left/right boundaries with depth — derived from sparse masks, robust across lighting and viewpoints, delivered on a real-time budget.

$ ./approach.sh

01Fine-tuned YOLO11 on a Roboflow road-segmentation dataset expanded 3× via augmentation for lighting and viewpoint robustness → 94% mAP50, 91.4% mAP50-95.

02Derived corridor geometry by extracting left/right road boundaries from sparse masks via distance transforms.

03Fused OAK-D W RGB-D depth for kart and bumper localization.

04Published the corridor to the planner as 3 Hz ROS2 topics — a stable contract between perception and planning.

$ cat results.json

94% mAP50 / 91.4% mAP50-95 road segmentation in varied lighting
Sparse masks → clean corridor geometry the planner consumes directly
Live on the kart — perception-to-planning pipeline running end-to-end
Led the perception sub-team; owned architecture and delivery

YOLO11 road segmentation — drivable region overlaid in green

Fig 1 — YOLO11 road segmentation on the kart's camera: drivable region overlay

The autonomous go-kart with ISC bodywork and roof-mounted perception rig

Fig 2 — the autonomous go-kart (ISC, UM-Dearborn) with roof-mounted perception

Close-up of the roof-mounted stereo depth camera and beacon

Fig 3 — the OAK-D W stereo depth camera on the perception mast

$ grep -r skills

team leadershipYOLO11semantic segmentationdata augmentationdistance transformsRGB-D sensor fusionOAK-D WROS2real-time perception