Abstract: In order to make a reasonable and effective prediction of vehicle's trajectory in the environment around the intelligent vehicle, a vehicle trajectory prediction method based on behavior recognition and curvature constraints is proposed. Firstly, it receives the perceived obstacle information and performs behavior recognition on vehicle combined with the lane line information provided by the high-precision map. Then the s-l coordinate system is established to decompose the vehicle motion into the motion along the lane line direction (longitudinal direction) and the motion perpendicular to the lane line direction (lateral direction). According to the behavior recognition result, the polynomial equation of the vehicle in the horizontal and vertical motion is obtained. Then the curvature of the lane line in the high-precision map is used as a constraint to select an optimal prediction trajectory. The actual vehicle experiment results show that under the three basic behaviors of lane keeping, lane changing and turning, the average vehicle trajectory prediction error within 4 s is 0.52, 0.51 and 1.03 m respectively, which is reduced by 1.81, 4.48 and 5.49 m compared with the prediction error of the CTRA model, and the average time of single vehicle trajectory prediction is 0.103 ms, which verifies the validity, accuracy and real-time of the proposed method.

Key words: vehicle trajectory prediction, behavior recognition, curvature constraints, high-precision map