Curve Trajectory Prediction through Vehicle Infrastructure Cooperation

The concept of vehicle infrastructure cooperative technology has been widely applied in the field of intelligent transportation. Through the technology, a novel vehicle trajectory prediction method is proposed in this paper, which can be used to provide a corresponding reference for risk prevention and control on the curve road. By integrating the multi-sensor data of the onboard and roadside sensors, the perception dimension is broadened and detailed vehicle information has been obtained. For the curve, the coordinates of the vehicles are transformed from the global coordinates to the Frenet coordinates based on the tangential and normal displacement of lane centre lines. A Bidirectional Long-Short Term Memory (B-LSTM) neural network is used to enhance the ability to process historical vehicle status information. Meanwhile, an encoder-decoder structure with the attention mechanism is applied to generate prediction trajectory points. A Mixture Density network gives a probability distribution for the final prediction output. In this paper, a curve scene is built by Unity to test the performance of the algorithm. Driving data is gathered by semi-physical simulation experiments from different drivers, after which a driver data set under the curve scene has been constructed. Utilizing vehicle infrastructure cooperative technology to expand the perception dimension, the trajectory obtained by the prediction algorithm has a smaller mean square error and higher accuracy in long-term prediction than relying on single on-vehicle sensors or on-road sensors. Through optimization of the curve, the prediction error has been reduced on longitudinal and lateral directions.