This project aims to improve the perception capabilities of autonomous vehicles by leveraging deep neural networks and recent advances in semi-supervised and self-supervised learning. Specifically, we seek to enhance key perception tasks such as road and lane detection, object detection, tracking, and trajectory forecasting, ultimately supporting more robust scene understanding and planning.
Traditional approaches to autonomous vehicle perception have heavily relied on supervised learning, requiring vast amounts of manually labeled data. While deep learning has led to significant improvements, scaling these models is increasingly challenging due to the cost and effort involved in data annotation. Additionally, perception systems must generalize across diverse driving environments, making it impractical to rely solely on supervised learning.
To address these challenges, this research will explore how semi-supervised and self-supervised learning can reduce the dependency on labeled data while maintaining or improving model performance. By leveraging large-scale unlabeled datasets, we aim to improve feature representations for key perception tasks, enabling more data-efficient and generalizable models.