Machine Learning-Based Prediction Models and Threat Detection for Lane-Keeping Assistance

Sammanfattning: Traffic accidents have been an ongoing problem for over a century and many efforts have been made to improve traffic safety. Historically, the focus has been on passive safety with innovations, such as crumpling zones, three-point seat belts, and airbags, that aim to mitigate the impact of collisions. As technology advanced, the focus shifted toward active safety, which aims to avoid accidents. Advanced driver assistance systems are nowadays utilized in vehicles to support the driver in critical situations where the driver is likely to fail the driving task. The system uses sensor information to estimate the risk of a threatful event, such as an unintended lane departure, and decides whether an automatic avoidance maneuver should be activated. However, from a legal perspective, it is the driver who is responsible for the driving, and consequently, the driver must be able to override an erroneous maneuver. This is an important aspect, as it restricts the system to the use of low-intensity maneuvers. That implies that a maneuver needs to be activated sufficiently early in time to be able to avoid the threatful situation, i.e., a long prediction horizon is needed to detect the threat in time. The decision to intervene with a supportive automatic avoidance maneuver is based on the output from the threat assessment, which uses a prediction model to estimate how the current traffic situation is evolving with time. Designing a well-functioning prediction model is challenging, as it must deal with multiple sources of uncertainties, such as sensor noise and drivers’ intentions, and it becomes even harder as the prediction horizon increases. This thesis focuses on how machine learning can be used to improve the performance of a lane-keeping assistance system. The goal has been to develop learning-based prediction models that are high-performing, robust, and efficient to compute in real time. The approach has been to evaluate the performance of linear and non-linear regression models using real-world data. The results show that both linear and non-linear prediction models are significantly better than a kinematic model. It also shows that linear prediction models are nearly as good as non-linear models, especially for shorter prediction horizons. However, the linear model is significantly easier to compute in real time and may therefore be a sufficient alternative for applications where computational power is restricted. Moreover, the robustness towards anomalies and samples that are out of the operational design domain can be improved by utilizing uncertainty-aware prediction models.