Abstract

A pedstrian trajectory prediction
method is proposed to address the problems of poor
accuracy in predicting pedestian motion trajectories
and high deviation in trajectory prediction when
pedestrians make significant turns in autonomous
vehicles. The method uses a LiDAR and an improved
trajectory prediction algorithm as the detection sensor.
By processing continuous single frame point cloud
data through streaming point cloud processing, the
minimum bounding box of pedestrians is clustered
and extracted, and its centroid is used as the
positioning coordinate as the input for the prediction
algorithm. At the same time, multiple motion models
are established based on the motion characteristics of
pedestrians, and an improved adaptive interactive
multi model unscented Kalman filtering algorithm is
used to predict pedestrian trajectories continuously.
The difference between the real and predicted
trajectories is used as a quantitative indicator to
compare the improved algorithm with the traditional
interactive multi model unscented Kalman filtering
algorithm. The experimental results show that the
adaptive interactive multi model unscented Kalman
filtering algorithm reduces the overall prediction
average error by 23.02% compared to the traditional
interactive multi model unscented Kalman filtering
algorithm and reduces the error peak value by 42.61%
during sudden pedestrian turns. It can effectively
predict the pedestrian's motion trajectory and has
better adaptability under sudden pedestrian turns.