Grid-Based Micro Traffic Prediction using Fully Convolutional Networks
| Autor: | Philip Schorner, J. Marius Zollner, Jonathan Hartl, Christian Hubschneider, Rupert Polley |
|---|---|
| Rok vydání: | 2019 |
| Předmět: |
Computer science
010401 analytical chemistry Monte Carlo method Process (computing) 020206 networking & telecommunications 02 engineering and technology computer.software_genre 01 natural sciences 0104 chemical sciences Overtaking 0202 electrical engineering electronic engineering information engineering Unsupervised learning Data mining Representation (mathematics) computer Dropout (neural networks) |
| Zdroj: | ITSC |
| DOI: | 10.1109/itsc.2019.8917263 |
| Popis: | We propose an approach for the model-free prediction of multi-lane traffic scenes using Fully Convolutional Networks (FCN). In order to generalize the applicability of the approach, we use a grid-based, sensor-agnostic input representation. Positioning, velocity, heading, turn indicator as well as information about occluded areas are stacked on the road layout creating a multi-layer input for the FCN. Training and evaluation data is acquired using a simulation that provides plausible vehicle movement for vehicle following, lane changing and overtaking. We demonstrate that the FCN architecture is able to implicitly learn the underlying vehicle kinematics, cooperative driving behavior as well as track vehicles through occluded areas. We further propose an iterative training process that outperforms single step predictions and is also suitable for unsupervised learning. The impact of loss representation, network depth and input resolution are evaluated to point out further improvements. As an extension, we evaluated the epistemic uncertainty obtained via Monte Carlo dropout in an approximation of a Bayesian Neural Network that also yields important information about prediction inaccuracy and possible multi-modalities. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|