Dynamic Interaction Graphs for Driver Activity Recognition

The drivers activities and the resulting distraction is relevant for all levels of vehicle automation. It is especially important for take-over scenarios in partially automated vehicles. To this end we investigate graph neuronal networks for pose based driver activity recognition. We focus on integrating additional input modalities like interior elements and objects and investigate how this data can be integrated in an activity recognition model. We test our approach on the Drive & Act dataset [1]. To this end we densely annotate and publish the bounding boxes of the dynamic objects contained in the dataset. Our results show that adding the additional input modalities boosts the recognition results of classes related to interior elements and objects by a large margin closing the gap to popular image based methods.