Hazard Detection in Driving Simulation using Deep Learning

The advancement in Big Data and Analytics has led to a data driven approach in research and development. Many traffic research studies use experimental simulators for data acquisition. This enables researchers to collect large amounts of data from human test subjects. Some research practices involve manual procedures to process participant data. On a large scale this is very labor intensive, highly time consuming and expensive. In these cases, automated solutions to perform these tasks can be implemented using deep learning techniques. This paper considers alternative approaches to process data from a longitudinal research study that uses a high-fidelity driving simulator with integrated eye tracking technology. In the simulation, specific roadway hazard scenarios are programmed to replicate realistic driving hazards. These hazard elements are then annotated by a researcher for behavioral analysis. This process is typically done manually on a frame-by-frame level with the help of software. The process is then repeated for every participant. It is a long and tedious task requiring long hours and high labor costs to be completed reliably. A deep learning solution is implemented to automate the hazard detection and annotation process. This method is tested across a small subset of participant data to assess the performance of this solution. With this method, the labor intensive process can be mitigated and improve the overall time and labor efficiency and to ensure timely dissemination of results from this research study.