Electronic Components and Systems for Automotive Applications by Unknown

Author:Unknown
Language: eng
Format: epub
ISBN: 9783030141561
Publisher: Springer International Publishing

One often used method to increase the detection performance is to fuse the raw data signal with neural networks that learn robust features from multiple inputs. In case of a disturbed sensor stream the fusion is not so easy anymore (or straight forward to do), because the variations and complexity of the disturbances are so great that they can never be fully registered and learned. Neural networks performing a Feature Level Sensor Fusion can be particularly error prone to disturbances outside the training data distribution, even if the other undisturbed sensor stream provide high-quality data (Bijelic and Münch 2018).

To overcome this problem, we propose a simple data augmentation scheme that ignores data from degraded sensors, although it has never seen these degraded characteristics of this sensor during training. This can be summarized as a learned “OR” operation at the fusion stage. This learned operation generally applies to all types of disorders that do not occur during training.

The learning scheme for increasing robustness is as follows (see also Bijelic and Münch 2018): We first train a dual-stream basic fusion network with a sufficiently large set of undisturbed data pairs. In order to make it insensitive to disturbances, we now train in a second step the network with data pairs where one element (stream) contains data of the background class and the other element (stream) contains data of the undisturbed object class. Already with a share of 10–20% of disturbed data pairs we generate a fusion network, which is robust against disturbances in a stream. In our experiments, the percentage of disturbed data pairs in the base network could be up to 50% without significantly changing the behavior of the fusion network (Fig. 13).

Fig. 13Proposed learning technique: Training a fusion network with mixed examples out of background and object class enables the network to learn a ‘or’ operation between the two input channels. This enables to retrieve information from one channel even if the second channel is disturbed with unknown disturbances

Download

Copyright Disclaimer:
This site does not store any files on its server. We only index and link to content provided by other sites. Please contact the content providers to delete copyright contents if any and email us, we'll remove relevant links or contents immediately.