Autonomous Car

Overview

The page has five panels. The first steps a camera frame through the lane detector with every calibration value on a slider. The second animates the node graph at its real message rates, with a topic echo pane and the startup probe. The third closes the loop around a simulated car with adjustable gains and strip charts. The fourth replays a garbage run through its state machine, servo duty cycles and command timeline. The fifth shows the real car in photographs, clips and a recorded lane-following run.

Lane vision

The lane node began from the course teaching assistant's example and was reworked for this car. It watches one band of each frame, keeps pixels whose HSV color falls in a tuned yellow range, and cleans the mask with a blur, one erosion and four dilations before the width filter. Only contours whose rotated bounding box is between 15 and 112 pixels wide count. Their centroids become a single steering error measured from a centre line at 55 percent of the frame width. With one dash in view the error points at it. With several, the node averages them on a straight and chooses the nearest line outside the threshold band on a curve.

A separate camera node runs a Roboflow garbage detector on the OAK-D Lite itself and publishes a detection flag, a steering error toward the target, and the width of the widest box in pixels.

ROS 2 pipeline

Four standing nodes share six topics. The camera node publishes the frame and the detector outputs, the lane node publishes its error on /centroid, and the guidance node turns whichever error it follows into /cmd_vel for the motor controller. A fifth node, the servo sweeper, is spawned only when the arm has to move. At startup the guidance node listens for detections for three seconds and then commits the whole run to either lane-following or garbage-seeking.

Control

Steering comes from a PID controller with gains of 0.2, 0 and 0.1, clamped to plus or minus one. Throttle is scheduled on the error: 0.2 inside the threshold band, easing linearly to 0.1 as the error grows, so the car slows into corners. When the widest detection passes 200 pixels the node stops, spawns the servo sweeper to swing the arm from 0 to 100 degrees over I2C, and after five seconds runs a fixed five-step maneuver before resuming.

The demo notes two properties of the controller as written. The derivative term divides by a fixed 1/20 second while frames arrive at 10 hertz, which doubles the effective derivative gain, and the integral term is clamped so tightly that the integral gain has no effect.

Development

The demo was built in September 2026 from the team's public repository. At build time the original Python nodes are imported under stand-in ROS, camera and servo modules, and their outputs become test fixtures for the TypeScript ports. Original ROS 2 nodes run at build time as fixtures for the TypeScript ports: the lane port reproduces every intermediate image bit for bit against OpenCV 4.11 across 12 frames and 8 parameter sets, and the guidance port sends the same 448 /cmd_vel messages over a 45-second run. The camera frames, detections and vehicle physics are simulated, and the page says so.

See also

• Computer Vision
• HardHack 2026
• David's Internet

References

1. ^ content/autonomous-car/README.md, David's Internet.
2. ^ 148-spring-2025-final-project-team-3, UCSD Silberman Classes and Projects, GitHub.