An Approach to Vehicle Trajectory Prediction Using Automatically Generated Traffic Maps

Trajectory and intention prediction of traffic participants is an important task in automated driving and crucial for safe interaction with the environment. In this paper, we present a new approach to vehicle trajectory prediction based on automatically generated maps containing statistical information about the behavior of traffic participants in a given area. These maps are generated based on trajectory observations using image processing and map matching techniques and contain all typical vehicle movements and probabilities in the considered area. Our prediction approach matches an observed trajectory to a behavior contained in the map and uses this information to generate a prediction. We evaluated our approach on a dataset containing over 14000 trajectories and found that it produces significantly more precise midterm predictions compared to motion model-based prediction approaches. I. MOTIVATION AND RELATED WORK In automated driving, planning and understanding the ego trajectory is one of the most important and fundamental tasks. It is necessary in order to plan emergency maneuvers, enable automatic lane keeping, dynamically adapt the velocity or in order to drive autonomously. To solve this task it is necessary to know how the car’s surroundings will change during the planning horizon. Since these changes cannot be measured or known in advance it is necessary to perform some form of prediction. For static surroundings or obstacles this task is trivial. However, for moving objects and other traffic participants in particular the estimation of future behavior becomes a very challenging problem. In commonly used motion planning tasks, the behavior of traffic participants is fully expressed by their trajectories. Therefore, we focus on trajectory prediction in this work. Whereas pedestrians and in many cases also cyclists move arbitrarily on free areas, roadways restrict vehicle motion to certain trajectory patterns. Thus, we target the identification and extraction of vehicle trajectory patterns in this work and present a framework to predict vehicle trajectories. Vehicle trajectory prediction is not an exact and deterministic problem since an observer does usually not have all the relevant information such as a driver’s intention or driving style. Many prediction approaches make the assumption that these factors cannot be reliably estimated and therefore only *The research leading to these results has received funding from the German collaborative research center “SPP 1835 Cooperative Interacting Automobiles” (CoInCar) granted by the German Research Foundation (DFG). 1Authors are with Institute of Measurement and Control Systems, Karlsruhe Institute of Technology, Karlsruhe, Germany {quehl,haohao.hu,martin.lauer}@kit.edu 2Author is with FZI Research Center for Information Technology, Karlsruhe, Germany. [email protected] Fig. 1: Comparison of different prediction approaches at two points in time. CYRA in orange, MRM-based prediction in red (solid and dashed) and our proposed method in green. use previously known information about vehicle dynamics and the current movement state of the observed vehicle for their trajectory prediction. For example, it is known that given a current velocity the other vehicle cannot accelerate or decelerate faster than the engine or breaks physically allow and that further the driver aims too keep some degree of comfort for all passengers. Therefore the future vehicle velocity and acceleration can be estimated within certain bounds. A similar assumption can be made about the curvature the vehicle can drive based on the maximum steering angle and vehicle stability. However, these bounds still do not limit the amount of trajectories to a size for which an accurate prediction is possible. Previous work by Schubert et al. [1] showed that assuming a Constant Yaw Rate and Acceleration (CYRA) model provides good results for vehicle tracking tasks. This suggests that for short-term prediction, e.g. the time between two vehicle detections, such a simple model based on vehicle dynamics provides good predictions. This was applied for example in [2] and [3] for short-term vehicle trajectory prediction. For long-term predictions on the other hand such predictions can become quite inaccurate since they cannot predict changes in the yaw rate that each person would be able to predict e.g. when leaving or entering a bend. A different approach was pursued in [4] and [5] which incorporated a Maneuver Recognition Module (MRM) in order to identify actions from a certain set of maneuvers and using that for trajectory prediction. Houenou et al. [6] combined motion model and maneuver ar X iv :1 80 2. 08 63 2v 1 [ cs .C V ] 2 3 Fe b 20 18 based predictions trying to benefit of both approaches. A disadvantage with maneuver based approaches, however, is that the maneuvers recognized by the MRM may be predictable a lot sooner. A MRM for example needs a short time in order to recognize that a car starts entering a bend even though any person that sees that a bend is coming could make a better prediction far sooner. This paper proposes a trajectory prediction method that uses a special graph based map representation generated through trajectory observations to predict future changes in maneuver and trajectory. By using information contained in a map this approach is able to predict driving maneuvers along the pathway of lanes and intersections and assign probabilities to different driving behavior. Our prediction method is facilitated by statistical information about natural driving behavior in the surrounding area. Figure 1 illustrates the different approaches with a simple example: Before the car starts turning neither an MRM nor a motion model based approach can make a correct prediction. In the next timestep an MRM-based method might yield a correct prediction, however if a wrong maneuver is estimated (lane change instead of turning), it might still make a wrong prediction. CYRA on the other hand will provide a good short-term but a bad long-term prediction either way. First, we provide an overview on our automatic map creation in Section II. Based on this map information, we present our novel prediction method in Section III. By evaluating our method in Section IV, we show that it is capable to make accurate mid-term predictions. Finally, we conclude our work in Section V.