Automotive Crash Reconstruction in Real-Time via Stateful Stream Processing

Vehicle telematics systems generate continuous streams of sensor data that can be leveraged for road accident analysis. Individual events in these streams do not represent complete crash events, as they contain only partial information about vehicle position, speed, and acceleration before and after an accident. Reconstructing a full crash scenario therefore requires aggregating and analyzing multiple related events. This thesis presents the re-engineering of a crash reconstruction service based on stateful stream processing. The service ingests inbound event streams, normalizes the data, and aggregates crash-related events into a unified crash entity. Once formed, the crash entity is used to estimate properties such as crash reliability and severity, enabling automated accident classification. The legacy reconstruction system was developed over a decade ago, using the tools available at the time. These tools were not designed for event-processing workloads and therefore do not effectively minimize input/output latency. In a system where thousands of events are processed every second, even small delays can have a significant impact. Increased processing latency requires running additional service instances to keep up with demand, ultimately raising the overall cost of the system. By leveraging stateful stream processing abstractions, the proposed approach achieves high throughput and scalability while reducing the complexity associated with traditional manual state management. The resulting system meets the performance requirements typical of large-scale vehicle telematics applications.