Short-term motion prediction of autonomous vehicles in complex environments: A Deep Learning approach

Dulian, Albert

Abstract

Complex environments manifest a high level of complexity and it is of critical importance that the safety systems embedded within autonomous vehicles (AVs) are able to accurately anticipate short-term future motion of agents in close proximity. This problem can be further understood as generating a sequence of coordinates describing the plausible future motion of the tracked agent. Number of recently proposed techniques that present satisfactory performance exploit the learning capabilities of novel deep learning (DL) architectures to tackle the discussed task. Nonetheless, there still exists a vast number of challenging issues that must be resolved to further advance capabilities of motion prediction models.
This thesis explores novel deep learning techniques within the area of short-term motion prediction of on-road participants, specifically other vehicles from a points of autonomous vehicles. First and foremost, various approaches in the literature demonstrate significant benefits of using a rasterised top-down image of the road to encode the context of tracked vehicle’s surroundings which generally encapsulates a large, global portion of the environment. This work on the other hand explores a use of local regions of the rasterised map to more explicitly focus on the encoding of the tracked vehicle’s state. The proposed technique demonstrates plausible results against several baseline models and in addition outperforms the same model that instead uses global maps. Next, the typical method for extracting features from rasterised maps involves employing one of the popular vision models (e.g. ResNet-50) that has been previously pre-trained on a distinct task such as image classification. Recently however, it has been demonstrated that this approach can be sub-optimal for tasks that strongly rely on precise localisation of features and it can be more advantageous to train the model from scratch directly on the task at hand. In contrast, the subsequent part of this thesis investigates an alternative method for processing and encoding of spatial data based on the capsule networks in order to eradicate several issues that standard vision models exhibit. Through several experiments it is established that the novel capsule based motion predictor that is trained from scratch is able to achieve competitive results against numerous popular vision models. Finally, the proposed model is further extended with the use of generative framework to account for the fact that the space of possible movements of the tracked vehicle is not strictly limited to single trajectory. More specifically, to account for the multi-modality of the problem a conditional variational auto-encoder (CVAE) is employed which enables to sample an arbitrary amount of diverse trajectories. The final model is examined against methods from literature on a publicly available dataset and as presented it significantly outperforms other models whilst drastically reducing the number of trainable parameters.

Citation

Dulian, . A. (2024). Short-term motion prediction of autonomous vehicles in complex environments: A Deep Learning approach. (Thesis). University of Hull. Retrieved from https://hull-repository.worktribe.com/output/4625808

• Thesis Type: Thesis
• Deposit Date: Apr 16, 2024
• Publicly Available Date: May 1, 2024
• Keywords: Computer science
• Public URL: https://hull-repository.worktribe.com/output/4625808
• Additional Information: School of Computer Science; University of Hull
• Award Date: Apr 9, 2024