@inproceedings {pub2930,
	title = {A novel approach to driver behavior prediction using scene context and physical evidence for intelligent Adaptive Cruise Control (i-ACC)},
	author = {Jens Schm{\"u}dderich AND Sven Rebhan AND Thomas H Weisswange AND Marcus Kleinehagenbrock AND Robert Kastner AND Morimichi Nishigaki AND Shunsuke Kusuhara},
	year = {2015},
	month = {September},
	abstract = {Conventional driver assistance systems react to another traffic participant{\textquoteright}s behavior as soon as it becomes apparent. In other words, they react as soon as the host vehicle{\textquoteright}s sensors detect the start of another vehicle{\textquoteright}s change of behavior. ACC systems, for example, react to a cutting-in vehicle when the sensors measure a significant lateral motion or displacement. However, a change of behavior is usually the effect of adapting to the current driving situation, i.e. a driver will change lane to overtake a slower vehicle driving ahead of him. Therefore attentive drivers are well capable of foreseeing or predicting another vehicle{\textquoteright}s behavior before this behavior actually starts. To bridge this gap between technical systems and human capabilities, future active safety technology should therefore base upon prediction to offer what we call {\textquotedblleft}assistance beyond sensing{\textquotedblright}.
In this contribution we will present a novel approach for the prediction of other vehicles{\textquoteright} behaviors. This approach combines a prediction based on the contextual situation of the predicted vehicle with physical evidence of its movement. The prediction based on the contextual situation, which we call context-based prediction, builds relations between vehicles and the infrastructure. These relations are evaluated by calculating features indicative for a future behavior. One indicator for example, evaluates if a vehicle is significantly faster than its predecessor and if there is a gap available on the neighboring lane into which the vehicle could cut-in. A suited combination of the resulting indicators leads to confidence scores for a set of future behavior alternatives. This kind of prediction explicitly excludes the phenomenological appearance of a behavior. That means, the prediction of a cut-in bases on the relations between the cutting-in vehicle and its surroundings, but it excludes any information about lateral displacement or a physical motion trajectory. As an effect, this kind of prediction approach is capable of predicting behaviors multiple seconds before they even start, but it requires a correct detection of the scene context.
The detection of the scene context poses an additional challenge for the sensor system. To achieve a reliable estimate of future behaviors in all cases, an additional physical prediction approach is used. This physical prediction compares a history of recently measured vehicle positions to a set of trajectories, each representative for a set of different behaviors. The pointwise comparison results in a likelihood estimate for each behavior from a set of different behavior options. This prediction approach is capable of reliably predicting a behavior as soon as it started and it does not require further context information. In this contribution we will show a novel combination of the context-based and physical prediction, characterized in keeping each prediction self-contained to trigger subtle host vehicle actions, but also combining the predictions to a most reliable, early prediction, which is used to trigger effective host vehicle actions for keeping the driver safe and comfortable.
In Honda{\textquoteright}s new intelligent Adaptive Cruise Control this prediction is used to trigger a two staged braking maneuver if a vehicle on the neighboring lane is predicted to cut-in: If only one of the prediction algorithms predicts a cut-in, a mild braking is applied, comparable to a driver releasing the gas pedal. By design, this is usually the case for early predictions calculated by the context-based prediction algorithm. As soon as both prediction algorithms predict a cut-in, the braking force is increased as required to stay safely behind the cutting-in vehicle.
In this contribution we will present a quantitative evaluation of the proposed prediction method. To this means, we extracted 15.000km from our recordings of driving with a prototype vehicle equipped with i-ACC on European highways. This subset is a representative sample in terms of varying traffic density, weather conditions, and driving style. To measure the proficiency of the proposed prediction system we evaluated two different aspects {\textendash} the prediction accuracy and the prediction horizon, i.e. the timespan before a cut-in for which an accurate prediction can be achieved.
The prediction accuracy is evaluated by Receiver-Operating-Characteristics (ROC). Therefore the prediction of cut-ins is considered as classification problem and a variable threshold is applied to the confidence values for a predicted cut-in behavior. If the prediction confidence exceeds the chosen threshold, the algorithm predicts a cut-in, otherwise it predicts no cut-in. The resulting prediction is compared to manually annotated ground-truth of cut-ins and leads to a true-positive-rate and a false-positive per hour score. By varying the threshold an optimal trade-off between true-positive-rate and false positives per hour is chosen. In this publication we chose a threshold leading to a true positive rate of 85\% of all targeted situations at one false positive in 10hours. It has to be noted, that in all these false-positive cases only one of the two prediction algorithms predicts a cut-in. Thus the resulting false-positive braking maneuvers are only mild and not strong.
To evaluate the prediction horizon we annotated the time of the cut-in by selecting the point in time when the right tire of the cutting-in vehicle touches the right lane marker. This point resembles the time when the best currently commercially available competitor system reacts to the cut-in. We measure the time between this annotated cut-in and the first positive prediction defined by the threshold chosen above. This evaluation shows that our system predicts a cut-in on average 2.4s before the annotation point. The maximum achieved prediction horizon is 7.2s before the annotation point. An analysis also indicates a dependence of the prediction time horizon and the prediction accuracy: Decreasing the threshold leads to an increased prediction time horizon and increases true positive rate, but at the cost of increasing false-positives.
To summarize this contribution presents a novel approach for predicting other vehicles behavior by combining a context-based prediction with a physical prediction. The evaluation shows the efficiency of this proposed approach, which is now commercially available in the 2015 Honda CR-V. A companion paper with the title {\textquotedblleft}Introduction of Intelligent Adaptive Cruise Control (i-ACC) {\textendash} a predictive safety system{\textquotedblright} details the system embedding and presents a subjective user test.
},
	publisher = {FISITA},
	booktitle = {Future Active Safety Technology Towards zero traffic accidents (FAST-zero)}
}