We have been developing a Smart Black Box (SBB) to augment traditional low-bandwidth data storage logging with value-driven high-bandwidth data capture. Data time histories associated with unexpected signals, high-risk, or anomalous conditions are stored in long-term memory and later uplinked to the cloud. We developed anomaly detection metrics for four typical hazards: object detection discrepancies, challenging lighting conditions, sensor measurement outliers and traffic collision ricks. A series of driving tests with raw data capture and an off-the-shelf automotive driver assistance system enables data analysis with respect to object identification accuracy and completeness.
The AnAn Accident Detection (A3D) dataset contains 1,500 video clips and manually labelled anomaly start/end times, and can be downloaded at: https://github.com/MoonBlvd/tad-IROS2019