Driving Risk Models for Predicting, Planning and Warning
Tim Puphal
DOI: https://doi.org/10.26083/TUPRINTS-00018933
Abstract:Automated cars and driver assistance systems constantly progress in complementing the human user in many parts of the driving task. Prominent examples include car-following on a highway, blind spot monitoring, recommending safe lane changes or even navigating on urban streets. This current trend has mostly originated due to affordable perception sensors and the improved speed of computer chips.
However, for a wider acceptance of self-driving cars, there is still a need to prove safety in terms of accidents and near-critical encounters caused by a technical system. Essentially, humans want technologies in which the reasons behind actions and warnings are known. This understanding helps trust to be increased and allows the driver to deliberately take over control from the system. The ultimate goal is to provide generic and transparent planning algorithms with considered safety margins.
In this dissertation, the presented challenge is tackled by developing analytical driving risk models and applying them to the relevant automotive domains of prediction, planning and warning. The models predict motion of vehicles along paths and incorporate several risk types, e.g., from collisions to sharp turns. Hereby, risks are composed of probabilities and severities and improve the behavior selection of the vehicle.
The dissertation is divided into three parts. Firstly, existing risk models of related work are enhanced with real-world uncertainties that arise from vehicle dynamics, unknown future environment changes and possible behavior alternatives of other vehicles. Analyses using accident data and normal traffic data show that this model has, amongst others, a higher fidelity than state-of-the-art time indicators. Secondly, a novel planning approach is introduced, which minimizes situational risks and maximizes utility and comfort to obtain ego velocity profiles. In all the statistical simulations of car-following and intersection driving, the approach successfully realizes a proactive maneuver. The major novelty of this planner is the intelligent inclusion of priorities between interacting vehicles. Lastly, the dissertation is concluded by leveraging risk-based planners for online driver warning with different car sensor setups and test locations, which shows their real-time applicability. Specifically, and in practice, the time predictions and low-risk trajectories are transformed into intuitive signal outputs for visualization to a driver.
To summarize, the proposed methods in this dissertation are based on fully transparent models with probabilistic formulations. This can be seen as a substantial contribution for the validation and advancement of intelligent robots; specifically, vehicles. Compared to simple reactive logics and data-driven machine learning methods, the approaches provide detailed information about the system’s situation understanding and reasoning for motion planning.
Even if they are not used as driving support technologies themselves,
they still could help to rate the driving proficiency and safety of other
existing platforms or, rather, the human driver. The basis is always formed by an integrated risk calculation that is parametrized from recorded car encounters and average variations in car dynamics. In this way, we may come a step closer to the goal of zero crashes with fewer traffic jams on roads and comfortable travel.
Engineering,Computer Science