Abstract: A trajectory prediction strategy based on traffic scene feature identification was proposed to improve the accuracy of dynamic vehicle trajectory prediction around autonomous cars in structured roads. The surrounding vehicles were detected, tracked and localized by lidars and integrated navigation system. The driving intention estimation model was used to judge the vehicle driving behavior and classify the traffic scenes. The vehicle motion trajectory was predicted based on vehicle motion model and Markov decision process for lane keeping and lane changing scenes, respectively. The results showed that the average prediction error based on the vehicle motion model is about 0.12 m for the lane keeping scene, when the prediction time domain is 4 s for the lane change scene, the average prediction error based on the Markov decision process is about 0.17 m when the prediction time domain is 8 s. Therefore, the strategy can accurately predict the trajectory of dynamic vehicles in real time.

Key words: autonomous cars, trajectory prediction, driving intention estimation, Markov decision processes