TY - GEN

T1 - Trajectory Modelling in Shared Spaces

T2 - 20th International Workshop on Multi-Agent-Based Simulation, MABS 2020

AU - Cheng, Hao

AU - Johora, Fatema T.

AU - Sester, Monika

AU - Müller, Jörg P.

N1 - Funding Information: Supported by the German Research Foundation (DFG) through the Research Training Group SocialCars (GRK 1931). The authors thank the participants of the DFG research project MODIS (DFG project #248905318) for providing data sets. F.T. Johora and H. Cheng—Contribute equally to this work.

PY - 2021/1/19

Y1 - 2021/1/19

N2 - Realistically modelling behaviour and interaction of heterogeneous road users (pedestrians and vehicles) in mixed-traffic zones (a.k.a. shared spaces) is challenging. The dynamic nature of the environment, heterogeneity of transport modes, and the absence of classical traffic rules make realistic microscopic traffic simulation hard problems. Existing multi-agent-based simulations of shared spaces largely use an expert-based approach, combining a symbolic (e.g. rule-based) modelling and reasoning paradigm (e.g. using BDI representations of beliefs and plans) with the hand-crafted encoding of the actual decision logic. More recently, deep learning (DL) models are largely used to derive and predict trajectories based on e.g. video data. In-depth studies comparing these two kinds of approaches are missing. In this work, we propose an expert-based model called GSFM that combines Social Force Model and Game theory and a DL model called LSTM-DBSCAN that manipulates Long Short-Term Memories and density-based clustering for multi-agent trajectory prediction. We create a common framework to run these two models in parallel to guarantee a fair comparison. Real-world mixed traffic data from shared spaces of different layout are used to calibrate/train and evaluate the models. The empirical results imply that both models can generate realistic predictions, but they differ in the way of handling collisions and mimicking heterogeneous behaviour. Via a thorough study, we draw the conclusion of their respective strengths and weaknesses.

AB - Realistically modelling behaviour and interaction of heterogeneous road users (pedestrians and vehicles) in mixed-traffic zones (a.k.a. shared spaces) is challenging. The dynamic nature of the environment, heterogeneity of transport modes, and the absence of classical traffic rules make realistic microscopic traffic simulation hard problems. Existing multi-agent-based simulations of shared spaces largely use an expert-based approach, combining a symbolic (e.g. rule-based) modelling and reasoning paradigm (e.g. using BDI representations of beliefs and plans) with the hand-crafted encoding of the actual decision logic. More recently, deep learning (DL) models are largely used to derive and predict trajectories based on e.g. video data. In-depth studies comparing these two kinds of approaches are missing. In this work, we propose an expert-based model called GSFM that combines Social Force Model and Game theory and a DL model called LSTM-DBSCAN that manipulates Long Short-Term Memories and density-based clustering for multi-agent trajectory prediction. We create a common framework to run these two models in parallel to guarantee a fair comparison. Real-world mixed traffic data from shared spaces of different layout are used to calibrate/train and evaluate the models. The empirical results imply that both models can generate realistic predictions, but they differ in the way of handling collisions and mimicking heterogeneous behaviour. Via a thorough study, we draw the conclusion of their respective strengths and weaknesses.

KW - Deep learning

KW - Game theory

KW - Mixed-traffic interaction

UR - http://www.scopus.com/inward/record.url?scp=85101310823&partnerID=8YFLogxK

U2 - 10.1007/978-3-030-66888-4_2

DO - 10.1007/978-3-030-66888-4_2

M3 - Conference contribution

SN - 9783030668877

T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)

SP - 13

EP - 27

BT - Multi-Agent-Based Simulation XXI

A2 - Swarup, Samarth

A2 - Savarimuthu, Bastin Tony

PB - Springer Science and Business Media Deutschland GmbH

CY - Cham

Y2 - 10 May 2020 through 10 May 2020

ER -