Adaptive lane tracking for autonomous vehicles using machine learning

• Main Author: Lee, Wei Tang
• Other Authors: Soong Boon Hee
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Engineering; Machine learning; Lane tracking
• Online Access: https://hdl.handle.net/10356/181573
• Institution: Nanyang Technological University

Autonomous vehicles are revolutionizing and they are the future of transportation. In order to let autonomous vehicles safe and efficient, they must be able to perceive and interpret information accurately from the surrounding, particularly in lane detection and lane tracking. However, the challenges of lane tracking lie in adapting to new and dynamic driving conditions, particularly where environmental factors like weather can drastically change. It is very difficult to maintain reliable lane tracking performance under such unpredictable and challenging circumstances. To address this issue, we aim to utilize the power of machine learning to develop an adaptive lane tracking network tailored for autonomous vehicle, specifically focusing on camera-based detection methods. In this project, we utilize a network that combines both Deep Convolutional Neural Network (DCNN) and Deep Recurrent Neural Network (DRNN) in the form of a UNetConvLSTM to perform lane tracking. This approach is beneficial as it leverages information from multiple frames, instead of just a single frame. This is to account for the connected nature of lanes and increase the lane prediction accuracy by utilizing more information provided by multiple frames used. With this advanced network, my adaptive lane tracking system aims to effectively adapt to challenging driving conditions where traditional methods often fail. To enhance the accuracy and efficiency of the lane tracking network, advanced techniques such as active learning and semi-supervised learning are explored. Active learning is used to select the most informative sample points based on uncertainty. This allows us to reduce the labelling effort required by just labelling the most informative samples. Additionally, an automated label generation algorithm is used to generate pseudo-labels for those informative unlabelled data. These pseudo-labels are then used to further train the model to enhance the model performance. This semi-supervised approach significantly improves model performance by expanding the training dataset without the need for extensive manual annotation, leading to a more robust and adaptive lane tracking system under varying driving conditions. The results of my adaptive lane tracking network demonstrate its high accuracy and reliability, even under varying driving conditions. In clear weather, the system achieved an impressive accuracy of around 98.09%, showcasing its ability to track lanes with great precision. Besides, my lane tracking network is also able to perform reliable lane tracking under more challenging scenarios, where traditional methods often struggle. For instance, when tracking lanes on curved roads, the model maintained a high accuracy of around 97.83%. When tracking lanes on road with shadows, which can confuse standard detection algorithms, the system still achieved a strong accuracy of approximately 97.69%. These results highlight the robustness and adaptability of our proposed approach, proving its capability to deliver reliable lane tracking even in challenging driving environments. In conclusion, my adaptive lane tracking network developed in this project has demonstrated its ability to effectively adapt to challenging driving conditions, including curved roads and roads with shadows, while maintaining high accuracy. The combination of advanced techniques such as active learning and semi-supervised learning has further enhanced the model’s performance, making it robust and reliable across different environments. Although the system performs well, there is room for improvement. Some potential enhancements and optimizations are discussed in the recommendations for future work section of this report, offering directions for further refinement of this lane tracking system.