Master of Science (MS)
End-to-End learning models trained with conditional imitation learning (CIL) have demonstrated their capabilities in driving autonomously in dynamic environments. The performance of such models however is limited as most of them fail to utilize the temporal information, which resides in a sequence of observations. In this work, we explore the use of temporal information with a recurrent network to improve driving performance. We propose a model that combines a pre-trained, deeper convolutional neural network to better capture image features with a long short-term memory network to better explore temporal information. Experimental results indicate that the proposed model achieves performance gain in several tasks in the CARLA benchmark, compared to the state-of-the-art models. In particular, comparing with other CIL-based models in the most challenging task, navigation in dynamic environments, we achieve a 96% success rate while other CIL-based models had 82-92% in training conditions; we also achieved 88% while other CIL-based models did 42-90% in the new town and new weather conditions. The subsequent ablation study also shows that all the major features of the proposed model are essential for improving performance. We, therefore, believe that this work contributes significantly towards safe, efficient, clean autonomous driving for future smart cities.
Guo, Dapeng, "END-TO-END LEARNING UTILIZING TEMPORAL INFORMATION FOR VISION- BASED AUTONOMOUS DRIVING" (2020). Master's Projects. 969.
Available for download on Wednesday, December 22, 2021