TY - JOUR TI - Inertial detection of unusual driving events for self-driving DO - https://doi.org/doi:10.7282/t3-ag2m-za37 PY - 2019 AB - While modern self-driving vehicles have shown their impressive capabilities towards offering new mobility to millions of people, it still remains challenging to build fully dependable and safe self-driving systems. To ensure the dependability of automated driving, the self-driving system is not only required to understand common road situations, but also widely different unusual events (e.g., objects on the roadway, pedestrian crossing highway, deer standing next to the road, etc.), which are very rare but more likely to cause unanticipated accidents. To detect unusual events, existing approaches seek to collect them by driving millions of miles with self-driving prototypes. But there still remains uncertainty because of limited miles covered. In contrast, this thesis proposes automatic unusual driving events identification algorithms, which can detect unusual cases through inertial sensing from in-vehicle devices in human-driven vehicles. This approach can be scaled to a much larger number of vehicles and thereby cover larger driving distances. The approach involves monitoring human driver reactions based on inertial sensors (e.g., accelerometer and gyroscope) and demonstrating that they are useful indicators of unusual driving events. Our inertial detection approach includes three stages. At first, we apply a potential emergency period (braking and swerving) detection to detect a sudden driver reaction, which reflects situations that challenge human drivers. Then we utilize three features to extract different properties from the detected periods. Finally, we propose a feature fusion method to fuse the features in an accuracy- driven way. Therefore, by extracting the features from the potential unusual periods, we can detect hazardous events from a large data set automatically. Besides, in order to improve the efficiency of processing a large scale of dataset, we also provide an alternative approach to process data in parallel pipelines in the cloud. We evaluate whether the inertially identified driving events match events that are manually labeled as unusual based on more than 120 hours of real-world driving data. The result shows the proposed fusion method outperforms the baseline methods with an 82% accuracy improvement for braking events as well as a 94% accuracy improvement for swerving events. To improve the dependability of automated driving systems, such detected events could be used in simulation tests to gradually refine self-driving software. KW - Electrical and Computer Engineering KW - Autonomous vehicles--Design LA - eng ER -