CV Pipeline for Self-Driving in BeamNG
Mar 2021 - May 2021
What?
Goal: Apply classical computer vision techniques to extract crucial data from vehicle-mounted cameras for autonomous driving.
Focus: Retrieve key information including lane presence, obstruction detection, and speed estimation from static images captured under specific conditions.
How?
Data Collection: Leveraged BeamNG.tech simulation, capturing images for processing through cameras that are virtually attached to the simulated vehicles.
Obstruction Detection: Used depth maps and OpenCV to detect obstructions and calculate minimum braking distance.
Lane Detection: Implemented lane detection using grayscaling, Gaussian Blur, Canny edge detector, and Hough Transform.
Speed Estimation: Employed the Lucas-Kanade method for optical flow estimation from image pairs.
Results
Obstruction Detection: Successful in detecting objects in front of the vehicle and determining whether to brake, despite challenges with image noise and depth map inaccuracies.
Lane Detection: Accurately detected lanes on the road, though susceptible to false positives due to environmental variations.
Speed Estimation: Provided consistent results, albeit with some errors compared to actual vehicle speed, indicating the method's limited reliability for real-time applications.
Check out the report below for more detail!