Experiments on vision-based simulation and real world lane following and obstacle avoidance with proximal policy optimization (PPO) | |
| Author | Min Khant Soe |
| Call Number | AIT Thesis no.DSAI-23-09 |
| Subject(s) | Automated vehicles--Control--Data processing Reinforcement learning |
| Note | A thesis submitted in partial fulfillment of the requirements for the degree of Master of Engineering in Data Science and Artificial Intelligence |
| Publisher | Asian Institute of Technology |
| Abstract | Lane following and obstacle avoidance are crucial tasks in autonomous driving. However, utilizing monocular vision for obstacle avoidance presents challenges due to the absence of 3D information required to estimate distances and lack capabilities in motion detection. Most fo the previous advancements in autonomous driving using deep reinforcement learning (DRL) have been limited to specific tasks such as lane following or obstacle avoidance in controlled environments. Furthermore, there is a lack of research exploring obstacle avoid ance solely using a front monocular camera mounted on a real car in real-world conditions. This thesis investigated the feasibility of using monocular vision for autonomous driving, with a special focus on the integration of Proximal Policy Optimization (PPO), a DRL tech nique, and monocular RGB vision. The aim is to establish monocular vision as a practical tool for autonomous navigation tasks, specifically lane following and obstacle avoidance. A novel approach is employed, where raw RGB images from a monocular camera are pro cessed through a UNet Semantic Segmentation model to produce semantic segmentation images. Enhancing the perception of motion, the study innovates by stacking three consec utive frames as input, providing a temporal perspective crucial for dynamic environments. This methodical approach incrementally introduces the agent to increasingly complex sce narios, enhancing learning efficiency and effectiveness. Additionally, the thesis explores various techniques for integrating control commands into the state input of the PPO agent, aiming to refine the learning process further. The PPO agent’s training, conducted in a simulated environment, emphasizes safety and feasibility, mitigating potential risks inherent in the learning process. While the agent demonstrated proficiency in the lane following task in both simulated and real-world conditions, it encountered challenges in the obstacle avoid ance task. The agent could navigate around obstacles effectively but struggled to realign with the ego lane post-avoidance, a natural maneuver for human drivers. Despite this limitation, the thesis represents a significant stride in autonomous driving research. It underscores the viability of monocular vision, combined with advanced DRL techniques such as PPO and innovative training methodologies, as a promising avenue for developing autonomous vehicles. This simplified yet effective approach offers a compelling alternative to more complex sensor systems. The findings of this study suggest that with continued refinement, monoc ular vision-based systems could evolve into a reliable, efficient choice for autonomous road navigation, ensuring safety and robust performance. |
| Year | 2023 |
| Type | Thesis |
| School | School of Engineering and Technology |
| Department | Department of Information and Communications Technologies (DICT) |
| Academic Program/FoS | Data Science and Artificial Intelligence (DSAI) |
| Chairperson(s) | Dailey, Matthew N.; |
| Examination Committee(s) | Mongkol Ekpanyapong;Chaklam Silpasuwanchai; |
| Scholarship Donor(s) | Asian Institute of Technology; |
| Degree | Thesis (M. Eng.) - Asian Institute of Technology, 2023 |