Potential of Driving Style Adaptation for a Maneuver Prediction System at Urban Intersections

Navigating through urban intersections is a challenging task for human drivers in gen-eral. More than 50 % of accidents with personal injuries caused by passenger car drivers in urban conditions happen at intersections. Currently, no driver assistance system in production is able to issue early warnings for pending collisions at urban intersections. One of the reasons is the warning dilemma at intersections arising from the variety of potential maneuvers a driver can perform. To overcome this dilemma, an approach is presented in this work to detect drivers' intentions on guidance level at urban intersections. The term guidance level is used here according to three-level hierarchy of the vehicle driving task described by Donges. The driver's core task on guidance level is to select a target track and target speed for safe driving. An intention here is understood as a driver's plan to execute a maneuver and is formulated before a maneuver is initialized. The goal of the intention detection system introduced here is to detect a maneuver intention based on data measured during inter-section approach and to predict a pending maneuver before commencement. The predic-tion quality achieved by the intention detection system solely based on automotive series sensors is analyzed. The turn indicator state is not used in the intention detection process at all. The approach introduced in this work is applicable for maneuver intention detection at arbitrary urban intersections. The intention detection is based on so-called “indicators”. Indicators have at least one input signal and one output signal for each potential intention. Indicators use transfer functions to calculate the maneuver likeli-hoods from any kind of input signal for all potential maneuvers. The indicators' transfer functions are calculated following the basics of Maximum-Likelihood principle. A quality measure to assess the quality of indicators and select indicators being beneficial for intention detection is introduced. The following data are used for intention detection: driver's control inputs during intersection approach, driver's head and gaze motion and mirror positions, environment perception information and intersection-specific information extracted from a digital map. Using independent indicators allows for the easy combination of different types of input signals in the prediction process. Different inference methods for the combination of indicators are discussed. Aside from inference methods with low computational complexity, a Bayesian network is applied as well. To analyze the feasibility of the approach introduced here, an experimental vehicle is equipped with a prototypic implementation of the intention detection system introduced in this work, using a close-to-production GNSS receiver and navigation map data for localization only. Test drives with 30 test subjects are carried out in the city of Darm-stadt. Data recorded in the test drives is used to train indicators' transfer functions and to evaluate the system's detection performance. A classification scheme for urban intersections is introduced and the performance evaluation is presented separately for different types of intersections. Due to inaccuracies arising from the localization, ego-motion based reference points are defined in this work for the system' s evaluation. These localization-inaccuracy-free reference points are calculated a posteriori to turn maneuvers based on the vehicle's motion. Using these reference points, average true-prediction rates of the implemented system on priority roads are 87.5 % for straight driving maneuvers, 81.5 % for right turns and 84.1 % for left turns, at 1 s second before maneuvers are initialized. All intention detection systems identified in related works focus on the detection of at least one intention connected with an action that the driver is about to perform. Here, a complimentary approach is introduced: The intention detection system is modified in order to exclude an intention related to a maneuver that the driver is not going to per-form. A number of exclusions are calculated for multiple horizons in front of the ego-vehicle. The advantage of this approach is that it overcomes the limitations of a classic "positive detection" approach: If a "positive detection" system cannot discriminate amongst at least two potential and concurring intentions, no decision is made. In this case the exclusion approach is able to exclude a third potential intention. Several studies can be found in literature addressing varying driving styles among dif-ferent drivers. Thus, this work analyses the potential of increasing true prediction rates by adapting the indicators' transfer functions to individual driving styles. All test sub-jects are classified into sporty, medium or relaxed drivers based on longitudinal acceler-ations tolerated by drivers in the study. The driving-style adaption process is able to increase prediction performance by more than 30 % for single maneuvers of drivers of the medium group at stop-sign- or give way-sign-regulated intersections. No appreciable effect to the detection performance could be found for other priority regulations