Detection of Evasive Maneuvers for Surrounding Vehicles

Abstract

Advanced Driver-Assistance Systems is a great aid to avoid traffic accidents, but today it relies on sensor sight. An evasive maneuver from a target vehicle is an example of a scenario where these systems have limitations. This thesis investigates the possibility to classify an evasive maneuver using a machine learning approach with a limited test-track dataset combined with a real-world dataset. Time series classification and data augmentation are used for different machine learning models for comparison. The result shows that the models are capable of predicting evasive maneuvers, however, it is clear by these results that further work to advance these models should be made. Data augmentation for time series data showed promising results in increasing the variety of input data and therefore increasing the limited dataset. A conclusion is made that a search algorithm that finds evasive maneuvers in larger datasets can be developed using the methods that are implemented in this thesis. Thus reducing data limitations that were an issue in this thesis. Future work could expand upon the models to include more complex features such as traffic interaction.