Out-of-distribution lane detector on a low-cost cyber-physical AV test bed

• Main Author: Gan, Shyan
• Other Authors: Arvind Easwaran
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Computer science and engineering::Computing methodologies::Image processing and computer vision
• Online Access: https://hdl.handle.net/10356/165946
• Institution: Nanyang Technological University

Safety and reliability are major challenges for machine learning systems deployed on safety critical real-time Cyber-Physical Systems (CPSs). One critical task is to monitor when testing data distribution shifts away from the distribution of training data, which may lead to erratic and unsafe outputs from the machine learning systems. Much progress has been made on Out-of-Distribution (OOD) detectors which are used to separate In-Distribution (ID) data from OOD data. Deep neural network based Variational Autoencoder (VAE) detectors are especially promising on CPS with limited computational resources which also require short inference times Computer vision based lane-following algorithms are well established in the field of autonomous vehicles. However, novel OOD lane markings could lead to unsafe steering inputs. In this project, a VAE is integrated into a lane following pipeline to perform OOD detection on lane markings. Different ways to integrate the detector are explored, and their performances compared. The resulting system is deployed on a Jetson Nano powered Duckiebot to show that our proposed detector can successfully stop the Duckiebot in the presence of previously unseen lane markings.