Vehicle trajectory prediction based on motion model and maneuver recognition
Summary (1 min read)
Introduction
- Active safety systems and self-driving cars are a promising solution to reduce the number of traffic accidents ([1], [2]).
- Predicting the trajectory of a vehicle is not a deterministic task since it depends on each driver’s intention and driving habits.
- It is based on the modeling and the comparison of the instantaneous path of the vehicle and the shape of the road.
- Then, a first trajectory prediction is made, only based on the recognized maneuver.
III. MANEUVER RECOGNITION MODULE (MRM)
- These can roughly be limited to these canonical cases: Keep lane θ and γ are the lane center line’s heading angle and its curvature in its closest point to the vehicle’s position.
- The distances Dk−i are calculated with Eq.8.
- If the distance is above the threshold, then the vehicle is either going to leave its current lane or has just entered it.
- For the sake of clarity, only change lane and keep lane maneuvers will be considered in the following.
IV. TRAJECTORY PREDICTION
- The method consists in mixing trajectory prediction based on maneuver recognition and trajectory prediction based on a motion model.
- Depending on the driver’s habits, the actual trajectory may be pretty smooth or pretty aggressive.
- The trajectories are first generated in the Frenet frame along the center line of the current lane of the vehicle (see Fig.3), then converted to the initial Cartesian coordinate system.
- This guaranties the jerk continuity and provides a unique solution.
V. EXPERIMENTAL RESULTS
- A prerecorded human driving data in semi-urban conditions was used to test the maneuver recognition algorithm and the trajectory prediction method.
- The integrated system recorded position data of the host vehicle with timestamps as explained in [16].
- Then, the lane change maneuver trajectories have been extracted from the recordings and conditioned to fit the same Cartesian coordinate system as depicted in Fig.6.
- The later are not considered in the following.
- Since the available data format is not rich enough to fit the requirements of the proposed method, the data set has been preprocessed in order to retrieve the missing information such as yaw angle, velocity, acceleration, yaw rate with their respective variances.
B. Trajectory prediction
- II shows the overall average values for all the extracted trajectories, per type of prediction.
- The accuracy of Tman does not vary a lot and the values show that the trajectory prediction based only on maneuver recognition is already close to the actual trajectory.
Did you find this useful? Give us your feedback
Citations
631 citations
410 citations
364 citations
250 citations
Cites methods from "Vehicle trajectory prediction based..."
...The invariable coefficient model integrates the physics- and maneuver-based models with a fixed coefficient function [8]....
[...]
...proposed a new trajectory prediction approach that combines a trajectory prediction based on the constant yaw rate and acceleration motion model and a maneuver recognition-based trajectory prediction [8]....
[...]
...In this experiment, the 40 lane-change cases are used to compute the average error of prediction position, and the average error of the invariable coefficient model [8] is studied and com-...
[...]
...Some methods integrate short- and long-term vehicle trajectory prediction approaches [8], [18]....
[...]
224 citations
References
567 citations
363 citations
"Vehicle trajectory prediction based..." refers background in this paper
...It appears that the model assuming Constant Yaw Rate and Acceleration (CYRA) gives the best results....
[...]
316 citations
"Vehicle trajectory prediction based..." refers background in this paper
...In [7] and [8], a set of trajectories corresponding to different maneuvers is predefined....
[...]
179 citations
"Vehicle trajectory prediction based..." refers background in this paper
...In [10] the ego vehicle trajectory is predicted by defining a driving context which is a vector containing relevant signals from sensors of the testbed such as light indicator or driver’s gaze direction....
[...]
165 citations
Related Papers (5)
Frequently Asked Questions (14)
Q2. What have the authors stated for future works in "Vehicle trajectory prediction based on motion model and maneuver recognition" ?
Future works include the estimation of the uncertainty along the predicted trajectories in order to estimate the TimeTo-Collision with an associated probability of collision for Collision Warning/Avoidance Systems.
Q3. What is the time interval used to define a unique trajectory?
The time interval ] 0, t(K) ] is then sampled and each sample time is used as maneuver ending time t1 to define a unique trajectory.
Q4. What was the purpose of the experiment?
A prerecorded human driving data in semi-urban conditions was used to test the maneuver recognition algorithm and the trajectory prediction method.
Q5. What is the jerk continuity of the lane?
d1 = d ∗ 1 ḋ1 = 0 d̈1 = 0 s̈1 = a0(11)For a change lane, d∗1 equals plus/minus the lane’s width depending on the direction of the maneuver and is null for a keep lane.
Q6. What is the simplest method for predicting a long term?
The method includes a prediction based on CYRA motion model which is very accurate for a short term and a prediction based on maneuver recognition which is more adapted for longer term prediction.
Q7. What is the lateral component of each trajectory?
The lateral component of each trajectory is of the form:d(t) = c5t 5 + c4t 4 + c3t 3 + c2t 2 + c1t+ c0 (12)Where ci,i={0,1,2,3,4,5} are coefficients.
Q8. What is the method for predicting a turn?
The method consists in mixing trajectory prediction based on maneuver recognition and trajectory prediction based on a motion model.
Q9. What is the lateral component of the trajectories?
The trajectories are then converted to the Cartesian coordinate system (see Appendix-B) and the best one is selected with respect to the cost function described hereafter.
Q10. What is the way to predict a turn?
based on the vehicle current state, the road parameters and the detected maneuver, a set of trajectories are first generated and the best one is selected with respect to a cost function described later.
Q11. What is the trajectories generated in the Frenet frame?
The trajectories are first generated in the Frenet frame along the center line of the current lane of the vehicle (see Fig.3), then converted to the initial Cartesian coordinate system.
Q12. What is the jerk continuity of the trajectories?
All the trajectories have the same initial state which is derived from the current state ζ0 = [x0, y0, θ0, v0, a0, ω0] of the vehicle in the Cartesian frame.
Q13. What is the jerk continuity of the vehicle?
Only the following assumptions are made: at the end state, the vehicle is moving right on the center line of its intended lane (known from the MRM) and has a constant longitudinal acceleration during the maneuver.
Q14. What is the average of the trajectories generated?
For real-time implementation, the complexity of the method can be kept low if the number of generated trajectories remains reasonable and if the curvature of the road is constant (in this case, the transformation from the Frenet frame to the Cartesian frame is trivial).