Feasible and adaptive attention-based models for multimodal trajectory prediction in urban driving scenarios
Loading...
URL
Journal Title
Journal ISSN
Volume Title
Perustieteiden korkeakoulu |
Master's thesis
Unless otherwise stated, all rights belong to the author. You may download, display and print this publication for Your own personal use. Commercial use is prohibited.
Authors
Date
Department
Major/Subject
Mcode
AUS
Degree programme
Language
en
Pages
68+2
Series
Abstract
A self-driving car which takes an autonomous decision needs three main building blocks, a perception module, a prediction module and a planning module. In this thesis, we consider the vehicle already capable of understanding the surrounding area; thus, we focus on the prediction module, which is responsible for predicting the future of the other agents in the scene. Thus, we examine, in particular, the prediction in urban driving scenarios in a multimodality setting where the model can learn to predict all the possible future scenarios in such complex environment. The predictions consist then of multiple sequences of coordinates plus the probability for each future. After having investigated the past and current methods, we have implemented different baselines, both deep learning methods and not. Hence, examining both the data that we used, and the network structures, we believe that some improvements are possible, and here we propose some methods to address those problems. First, we extend the previous loss with an additional term called offroad loss that penalise the model when the prediction lays outside of the road structure. Second, considering also that difficult scenes are rarer than simple scenes, we propose two different weighted sampling methods to overcome such imbalance, in this way, the model can adapt the prediction to more complicated and rare scenes. Finally, we try to extract more useful information from road structure, nearby agents and past information implementing different attention architectures inside the models. In this thesis, we also conduct a performance comparison between our methods and the baselines applying commonly used metrics. Moreover, to visually understand the impact of each method, we propose some anecdotal analysis showing the real differences in terms of prediction in some challenging situations.Description
Supervisor
Charambolous, ThemistoklisThesis advisor
Tran, Tuan AnhBerkemeyer, Hendrik