Multi-pedestrian trajectory prediction method based on multi-view 3D simulation video learning

Multi-pedestrian trajectory prediction is one of the key factors in integrating urban geographic information system and intelligent transportation. To address the problems of insufficient training data, difficult labeling, and low accuracy of pedestrian trajectory prediction in multi-view scenes for existing methods, we propose a novel multi-pedestrian trajectory prediction method based on multi-view 3D simulation video learning. First, a simulation simulator is used to generate the required multi-view pedestrian trajectory annotation data. Then, we mix up the trajectory of the selected view and the adversarial trajectory by a convex combination function to generate the enhanced adversarial trajectory. Next, an advanced detection and tracking algorithm is used to encode and track pedestrian appearance information. Furthermore, the enhanced trajectory and coding information are used as the feature input of a graph attention recurrent neural network to model pedestrian interaction. Finally, the pedestrian trajectory is decoded by a position decoder to extract pedestrian motion characteristics, and multi-pedestrian trajectory prediction is completed. The ADE and FDE accuracies of our method on the ETH/UCY fixed-view dataset are 0.41 and 0.82, respectively. The ADE accuracy on the ActEV/VIRAT and Argoverse multi-view datasets is 17.74 and 65.4, and the FDE accuracy is 34.96 and 172.8. © 2023 SinoMaps Press. All rights reserved.