Autoshuttle: A Novel Dataset for Advancing Autonomous Driving in Shuttle-Specific Environments

Autonomous shuttles transport passengers or goods without a human driver. For safety reasons and to achieve robust autonomy, they are usually equipped with advanced technologies such as artificial intelligence to detect potential hazards. However, existing datasets for autonomous driving are predominantly recorded in large cities, while shuttles usually operate within defined areas. Exisiting datasets do not cover all the classes required for our use case. The domain shift (position of sensors, background, and weather conditions) between exisiting datasets and our use case requires the need to record new dataset. Motivated by this observation, we create a new shuttle-specific dataset with multiple classes for deep learning-based object detection algorithms. We define the classes and recording scenarios by analyzing the shuttle route and selecting frequently occuring obstacles. Furthermore, we select data with a framerate of 2 Hz for getting it annotated with 2D and 3D bounding boxes. We anticipate that by recording shuttle specific dataset with use case defined classes and attributes, the trained object detection models will perform with improved accuracy in shuttle-specific scenarios. Ultimately, this dataset will enhance autonomy in different weather conditions and to estimate pedestrian behavior.