METHOD AND DEVICE FOR PREDICTING THE TRAJECTORY OF A TRAFFIC PARTICIPANT, AND SENSOR SYSTEM

Abstract

A computer-implemented method for predicting a trajectory of a traffic participant. Sensor data acquired by at least one vehicle sensor at a plurality of acquisition times is received. Based on the received sensor data, values of at least one motion parameter of the traffic participant are determined for each acquisition time. The trajectory of the traffic participant is predicted using a stochastic regression algorithm which receives the determined values of the at least one motion parameter of the traffic participant as an input.

Claims

1. A computer-implemented method for predicting a trajectory of a traffic participant, comprising the following steps: receiving sensor data acquired by at least one vehicle sensor at a plurality of acquisition times; determining, based on the received sensor data, values of at least one motion parameter of the traffic participant for each acquisition time; and predicting the trajectory of the traffic participant using a stochastic regression algorithm which receives the determined values of the at least one motion parameter of the traffic participant as an input.

2. The method according to claim 1, wherein the predicting of the trajectory of the traffic participant includes computing a predicted acceleration of the traffic participant using the stochastic regression algorithm, and integrating the predicted acceleration to compute the predicted trajectory of the traffic participant.

3. The method according to claim 1, wherein the stochastic regression algorithm is a Gaussian regression algorithm.

4. The method according to claim 1, wherein the determined values of the at least one motion parameter of the traffic participant include values for a position of the traffic participant, a velocity of the traffic participant, and an acceleration of the traffic participant, for each acquisition time.

5. The method according to claim 1, wherein an uncertainty of the predicted trajectory of the traffic participant is computed using the stochastic regression algorithm.

6. The method according to claim 5, wherein the predicted trajectory of the traffic participant and the uncertainty of the predicted trajectory of the traffic participant are used to calculate a probability of an accident and/or estimate a road topology.

7. The method according to claim 1, wherein trajectories of a plurality of traffic participants are predicted, wherein possible interactions between the plurality of traffic participants are taken into account for predicting the trajectories of the plurality of traffic participants.

8. The method according to claim 1, wherein the sensor data includes at least one of camera data or radar data.

9. A device for predicting a trajectory of a traffic participant, comprising: an interface adapted to receive sensor data acquired by at least one vehicle sensor at a plurality of acquisition times; a memory adapted to store the received sensor data; and a computer adapted to determine, based on the received sensor data, values of at least one motion parameter of the traffic participant for each acquisition time, and to predict the trajectory of the traffic participant using a stochastic regression algorithm which receives the determined values of the at least one motion parameter of the traffic participant as an input.

10. A sensor system for a vehicle, comprising: at least one sensor adapted to acquire sensor data; and a device for predicting a trajectory of a traffic participant based on the acquired sensor data, the device including: an interface adapted to receive the acquired sensor data at a plurality of acquisition times; a memory adapted to store the received sensor data; and a computer adapted to determine, based on the received sensor data, values of at least one motion parameter of the traffic participant for each acquisition time, and to predict the trajectory of the traffic participant using a stochastic regression algorithm which receives the determined values of the at least one motion parameter of the traffic participant as an input.

Description

BRIEF DESCRIPTION OF THE DRAWINGS

[0025] FIG. 1 shows a schematic block diagram of a sensor system of a vehicle according to an example embodiment of the present invention.

[0026] FIG. 2 shows an exemplary traffic scenario for illustrating an example method for predicting a trajectory of a traffic participant in the surrounding of a vehicle in accordance with the present invention.

[0027] FIG. 3 shows a schematic flow diagram of a method for predicting a trajectory of a traffic participant according to an example embodiment of the present invention.

[0028] FIG. 4 shows an exemplary average error of estimation in meters for the method according to FIG. 3 and a constant-turn-rate model.

[0029] FIG. 5 shows an exemplary error estimation of a standard deviation of a ratio of a real error and an estimated standard deviation of predictions for the method according to FIG. 3 and a constant-turn-rate model.

[0030] In the figures, like reference numerals designate corresponding similar parts.

DETAILED DESCRIPTION OF EXAMPLE EMBODIMENTS

[0031] FIG. 1 shows a schematic block diagram of a sensor system 1 of a vehicle. The sensor system 1 comprises at least one sensor 3, which may comprise a camera sensor, a radar sensor, a lidar sensor, an infrared sensor or the like. The sensor 3 acquires sensor data at each of a plurality of acquisition times. The sensor data may comprise camera images, radar images, and the like.

[0032] The sensor system 1 further comprises a device 2 for predicting a trajectory of a traffic participant. The traffic participant can be the vehicle itself. The traffic participant can also be a further vehicle, a pedestrian, cyclist, or the like. It is also possible to predict both the trajectory of the vehicle itself and of other traffic participants. The device 2 comprises an interface 21 coupled to the at least one sensor 3 and adapted to receive sensor data acquired by the at least one sensor 3. The device 2 further comprises a memory 22 which stores the received sensor data. The memory 22 may comprise a volatile or non-volatile data memory, e.g. a solid-state disk, memory card or the like.

[0033] A computer 23 of the device 2 is connected to the memory 22 and has access to the memory 22. The computer 23 may comprise at least one of a central processing unit (CPU), graphics processing unit (GPU), microcontroller, integrated circuit (IC), application-specific integrated circuit (ASIC), or the like.

[0034] The computer 23 analyzes the received sensor data to compute values of motion parameters of the traffic participant for each acquisition time. For instance, the computer 23 may determine values for a position, a velocity and an acceleration of the traffic participant for each acquisition time. The values of the motion parameters of the traffic participants are input for a stochastic regression algorithm. The position, velocity and acceleration may be given as two-dimensional vectors relative to a driving plane or can be three-dimensional values.

[0035] The stochastic regression algorithm is based on a stochastic process, i.e. a collection of random variables indexed by time, the random variables corresponding to the motion parameters of the traffic participant. Preferably, the stochastic regression algorithm is a Gaussian regression algorithm, also known as kriging. The Gaussian regression algorithm serves as a non-linear multivariate interpolation algorithm. Using the stochastic regression algorithm, the computer 23 computes predicted values of motion parameters of the traffic participant. In particular, the computer 23 may determine the predicted acceleration of the traffic participant for a plurality of time points in the future or as a continuous function of time. The components of the acceleration vector can be modeled as Gaussian stochastic processes, making it possible to handle prediction in an optimal, mathematically correct way. The method automatically adapts to the current situation by considering only the most recent measurements and estimating the probability distribution of the acceleration.

[0036] The computer 23 further integrates the distributions of the predicted acceleration of the traffic participant, using the determined position and velocity of the traffic participant at the acquisition times as initial values to determine integration constants. The computer 23 outputs the predicted trajectory, e.g. to a driver assistance system 6.

[0037] The computer 23 may further provide an uncertainty of the predicted trajectory of the traffic participant using the stochastic regression algorithm. The computer 23 may use the uncertainty, i.e. the error estimate, together with the predicted trajectory to calculate the probability of an accident for the vehicle with the traffic participant. The computer 23 may also estimate a road topology of a road the vehicle is driving on, using the uncertainty and the predicted trajectory.

[0038] The described prediction of the trajectory of the traffic participant can be carried out for a plurality of traffic participants in a surrounding of the vehicle. All trajectories can be predicted simultaneously by taking possible interactions between the traffic participants into account. For instance, the possible trajectories of the traffic participants can be reduced by excluding trajectories that would result in accidents between traffic participants and other traffic participants or the vehicle.

[0039] The driver assistance system 6 may control the vehicle based on the predicted trajectory of the further traffic participant. The driver assistance system 6 may be configured to carry out autonomous driving functions by controlling an acceleration, turn rate and the like of the vehicle.

[0040] Instead of predicting trajectories of further traffic participants, the device may also predict the trajectory of the vehicle itself. In this case, the sensor 3 may comprise an inertial sensor for measuring the acceleration of the vehicle.

[0041] FIG. 2 shows an exemplary traffic scenario for illustrating the method for predicting a trajectory T of a traffic participant 5 in the surrounding of a vehicle 4. Based on sensor data acquired by a sensor 3 of the vehicle 4, the computer 23 determines position x, velocity v and acceleration a of the traffic participant 5 for each acquisition time. The trajectory T the traffic participant 5 is driving on is predicted for future time points.

[0042] FIG. 3 shows a schematic flow diagram of a method for predicting a trajectory of a traffic participant.

[0043] In a first method step S1, at least one sensor 3 of the vehicle 4 acquires sensor data for each of a plurality of acquisition times. An interface 21 receives the acquired sensor data.

[0044] In a second method step S2, a computer 23 determines, based on the received sensor data, values of at least one motion parameter of the traffic participant for each acquisition time. The motion parameters may comprise a position, a velocity and an acceleration of the traffic participant. The motion parameters may also comprise a turn rate or the like.

[0045] In a further method step S3, the trajectory of the traffic participant is predicted. The method step S3 comprises a first sub-step S31, wherein the computer determines a predicted acceleration of the traffic participant using a stochastic regression algorithm. In a further sub-step S32, the computer 23 integrates the predicted acceleration to compute the predicted trajectory of the traffic participant 5.

[0046] FIG. 4 relates to the quality of the estimated trajectory. FIG. 4 shows an exemplary average of the square root of the squared error of estimation (E) in meters (m). The difference between the estimated trajectory from the actual trajectory (based on real measurements) is illustrated at some characteristic distances (d) in meters (m). The error of estimation is displayed for a reference model (M1) and for the method according to the invention (M2). The reference model is a constant-turn-rate model, i.e. assumes that the turn rate does not change within the time span under consideration. It can be seen that at short ranges at about 10 meters ahead of the vehicle, the error of the estimation decreases by about 75 percent as compared to the reference model M1. At middle ranges of about 50 meters ahead of the vehicle, the error of the estimation still decreases by about 50 percent. Also at larger distances of about 100 meters the difference is significant.

[0047] FIG. 5 shows an exemplary error estimation of a standard deviation of a ratio of a real error and an estimated standard deviation of predictions for the method according to the invention (M2) and the constant-turn-rate model (M1). The functions depend on distances (d) in meters (m). A target line A is shown, corresponding to a constant value of 1. FIG. 5 relates to a quality of the error estimation. Error estimation is a statistical property and can only be verified based on statistical quantities. The evaluation of the error estimation is a complex task, as both the real error and the estimated error are changing with elapsed time and distance. A real error of the i-th prediction is denoted by ϵ_i and the estimated standard deviation of the i-th prediction is denoted by σ_i. Multiple samples of pairs of real errors ϵ_i and estimated errors σ_i are examined. If the values of the estimated errors σ_i are good estimations for the values of the real errors ϵ_i, the standard deviation of ϵ_i/σ_i, i.e. Stddev(ϵ_i/σ_i), stays close to 1. This follows from the fact that each of the ϵ_i samples will be normalized with the σ_i estimated error. Thus, if the estimated error is statistically less than the real error, i.e. the model is too optimistic, the standard deviation of the ratio will be greater than 1. If the estimated error on the other hand is statistically greater than the real error, i.e. the model is too pessimistic, the standard deviation of the ratio will be smaller than 1.

[0048] As can be seen from FIG. 5, the reference model M1 is too optimistic about its prediction, the errors it really makes being much larger than it assumes. Based on the new approach, the error estimation is reliable as it is close to the target of 1.

[0049] The reliability of the error estimation makes it possible to use the approach in combination with other models. The approach is neither too optimistic nor too pessimistic about its own uncertainty. As a result, during the fusion of the approach with other trajectory estimation methods, the correct trajectory prediction uncertainty will not interfere in the trajectory fusion algorithms.