Improved Benefits of Collision Avoidance by Steering Technology on Real Life Safety, part II

Purpose and goal

Our focus has been to study the use of machine learning with the aim to study its ability to predict unintentional lane changes by processing time series data. We have focused on implementation aspects, with the goal of developing an efficient prediction model with low computational complexity, which has comparable or better performance than today´s traditional methods.

Expected results and effects

The results describe how to use machine learning and time series data in a computationally efficient way to achieve performance that surpasses the kinematic models that currently dominate the industrial applications. One of the most important results is that a computationally effective multi-step prediction model based on linear regression proved to provide unmatched performance at a low cost. Non-linear models worked better but only for relatively long prediction horizons.

Planned approach and implementation

The overall research method has been to implement and train new algorithms on existing, representative, and comprehensive data collected in real traffic, and to constantly evaluate and compare with the performance of well-known existing methods. To further increase the value of the results, many of existing alternative methods have often also been implanted to get a more comprehensive understanding of the results. The project has been run as an industrial PhD-project in collaboration with senior researchers at Zenseact and Chalmers Tekniska Högskola.