With the rapid development of autonomous driving technology, accurate vehicle trajectory prediction has become a key challenge. This article focuses on vehicle trajectory prediction in automatic driving system, aiming at building an efficient MMDR model to improve the accuracy and stability of trajectory prediction. In this article, based on the strategy of feature level fusion, the features of different modal data are spliced and fused to construct a hierarchical DNN (Deep neural network) structure, and the algorithm is optimized by using attention mechanism and other technologies. The experimental results show that the MMDR model (Multi-modal depth reasoning algorithm model) proposed in this article is obviously lower than the single modal model in MSE (Mean square error) index. At each test sample point, the MSE value of the single modal model such as CNN (Convolutional neural network) fluctuates greatly, while the MMDR model is more stable and has a lower value. In mIoU (Mean Intersection over Union), MMDR model achieves about 0.85 in urban road scene, keeps stable above 0.9 in expressway scene, and keeps around 0.8 in intersection scene, which is significantly better than other comparison models. It can be seen that the model performs well in the accuracy and stability of vehicle trajectory prediction, and can effectively meet the trajectory prediction requirements of automatic driving system in complex scenes.
Kaixi Zhang (Sun,) studied this question.