Adaptive self-supervised model for trajectory prediction

Abstract

The prediction of trajectory holds significant importance in improving vehicle safety and enhancing traffic intelligence. Since the existing models ignore the problem of trajectory bias due to possible missing sensor data in realistic scenarios, this paper presents a novel approach to enhance the accuracy of trajectory prediction utilizing a CNN-Transformer adaptive complementary self-supervised model. The model combines complementary learning of CNN local features and Transformer global features, building upon the advancements of the PishguVe model. Furthermore, the present study employs masks to simulate sensor data aberrations or data loss that may occur during real driving scenarios. Additionally, self-supervised learning is utilized to enhance the robustness and generalization of the model. Upon evaluation of the proposed model on the CHD eye-level datasets, it is observed that the Average Displacement Error (ADE) decreased to 18.02, which is 7.25% lower than that of the current leading model, the PishguVe model. When assessing the CHD high-angle datasets, the Average Displacement Error (ADE) is determined to be 16.80 pixels, and the Final Displacement Error (FDE) is calculated to be 59.67 pixels. The algorithms in question exhibit lower values compared to the currently optimal PishguVe's ADE metric of 3.28% and FDE metric of 3.74%, demonstrating the superior performance of the proposed model.