Reviewing Deep Learning-Based Feature Extractors in a Novel Automotive SLAM Framework

Simultaneous Localization and Mapping (SLAM), which is characterized as a core problem in autonomous vehicles, involves the estimation of the vehicle's position and the concurrent building of the map of the environment. The use of deep learning-based feature extractors has gain increasing popularity since they possess the ability to extract reliable and repeatable features from raw sensor data. However, the performance of deep learning-based approaches varies depending on the application, environmental conditions, and the type of implemented technology. In this paper, we evaluate the performance of several deep learning-based feature extractors integrated into a SLAM system, using as input real and synthetic data, which implement common odometry problems. To our knowledge, this is the first work that benchmarks the accuracy of deep-learning based algorithms in estimating the vehicle's trajectory in specific odometry corner cases.