Toward Autonomous Driving: Data-Model Integration for Trajectory Reconstruction and Driving Action Extraction

With the development of autonomous driving, understanding the behavior of traffic participants has become increasingly important. Vehicle trajectories contain rich information, however, they are often contaminated with noise in real-world environments. While existing trajectory processing methods may incorporate physical models to some extent, they typically struggle to enforce strict adherence to physical constraints. As a result, the processed trajectories may become distorted, making it difficult to accurately infer the driver’s control actions. In this paper, the vehicle kinematic model is embedded into a learnable neural network. By leveraging backpropagation through the physical model, the network can be trained without the need for explicit labels of driving actions. The design of intermediate network layers ensures that the output trajectories strictly satisfy physical constraints and enables robust inference of the driver’s inputs. Cross-dataset experiments demonstrate that the proposed method significantly outperforms direct computation approaches: errors in acceleration and deceleration estimation are reduced by approximately 84%, while steering action errors are reduced by approximately 43%. Real-world vehicle data further validate the trajectory reconstruction method, achieving a 5% reduction in prediction error when applied to fixed-point trajectory prediction models compared to unprocessed trajectories.