Deep Learning enables weather independent Re-Localization
Direct visual-SLAM (Simultaneous localization and mapping) methods have shown exceptional tracking performance. However, they still suffer from lighting and weather changes. In particular, re-localization in a pre-built map is a challenging problem since the visual appearance of the environment is changing over time. Recently, Artisense researchers came up with a novel solution to overcome these limitations and give answers to some of these challenging questions.
Artisense: For our Re-localization approach weather conditions don’t matter anymore.
With its deep learning (AI) based approach Artisense’ re-localization system gets unaffected by lighting and weather changes. The key to achieve this is by teaching a neural network to learn that the same scene recorded under different conditions should result in a similar representation. In practice, the network learns a higher-dimensional abstraction of the image which can be used for tracking. The learned representations will be robust to changes in lighting and weather compared to normal images.
A real-world example would be the recording of one route in both, summer as well as in winter conditions. In both cases, similar image abstractions are created which allow a precise alignment with respect to the map. The following video shows how re-localization works in a robust way under any weather conditions. This is especially required for drift-free localization in GNSS-denied environments. Use case for self-driving cars is localization in e.g. tunnels or urban canyons.
In a future generation, the Artisense re-localization engine can be integrated into any existing HD map.
You’ll find more detailed information in the regarding paper here.
VIDEO: It demonstrates how Artisense’ centimeter-accurate vision-based re-localization solution can be used to reliably localize in a pre-built map. This real-time system is robust against different lighting and weather conditions and only uses off-the-shelf cameras.