| 1. |
- Arnelid, Henrik, et al.
(författare)
-
Recurrent Conditional Generative Adversarial Networks forAutonomous Driving Sensor Modelling
- 2019
-
Konferensbidrag (refereegranskat)abstract
- Simulation of the real world is a widely researchedtopic in various fields. The automotive industry in particular isvery dependent on real world simulations, since these simulations are needed in order to prove the safety of advance driverassistance systems (ADAS) and autonomous driving (AD). Inthis paper we propose a deep learning based model for simulating the outputs from production sensors used in autonomousvehicles. We introduce an improved Recurrent ConditionalGenerative Adversarial Network (RC-GAN) consisting of Recurrent Neural Networks (RNNs) that use Long Short-TermMemory (LSTM) in both the generator and the discriminatornetworks in order to generate production sensor errors thatexhibit long-term temporal correlations. The network is trainedin a sequence-to-sequence fashion where we condition theoutput from the model on sequences describing the surroundingenvironment. This enables the model to capture spatial andtemporal dependencies, and the model is used to generatesynthetic time series describing the errors in a productionsensor which can be used for more realistic simulations. Themodel is trained on a data set collected from real roads withvarious traffic settings, and yields significantly better results ascompared to previous works.
|
|
| 2. |
|
|
| 3. |
- Zec, Edvin Listo, et al.
(författare)
-
Recurrent Conditional GANsfor Time Series Sensor Modelling
- 2019
-
Konferensbidrag (refereegranskat)abstract
- Simulation of the real world is a widely researchedtopic in many different fields, and theautomotive industry in particular is very dependenton real world simulations. These simulationsare needed in order to prove the safety ofadvance driver assistance systems (ADAS) and autonomousdriving (AD). In this paper we proposea deep learning based model for generating timeseries outputs from sensors used in autonomousvehicles. We implement a Recurrent ConditionalGenerative Adversarial Network (RC-GAN) consistingof Recurrent Neural Networks that useLSTMs in both the generator and the discriminatorin order to generate sensor errors described astime series that exhibit long-term temporal correlations.The network is trained in a sequence-tosequencefashion where we condition the modeloutput with time series describing the environment,which enables the model to capture spatialand temporal dependencies. The RC-GAN is usedto generate time series describing the errors in aproduction sensor on a data set collected fromreal roads, and yields significantly better resultsas compared to previous works on sensor modelling.
|
|
