Study on Sparse Vehicle Trajectory Reconstruction Based on Fully Connected Neural Network

Addressing the challenge of reconstructing vehicle trajectories in areas with sparse electronic police (E-police) coverage, this study proposes a novel algorithm for trajectory reconstruction utilizing license-plate-recognition (LPR) data. After removing invalid plates, single detections, and taxi or ridehailing records, vehicle trajectories are extracted. Road speeds are calibrated using GPS data from taxis. Trajectories are segmented into origin-destination (O-D) pairs based on calibrated speed limits and minimum travel time. The K-shortest path algorithm generates candidate trajectories that minimize travel time. Six indicators - the distance, time difference, number of signalized intersections, number of left turns, number of right turns, and road grade of the candidate trajectories are calculated and normalized with the Min-Max scaler. A fully connected neural network model is trained on these multi-dimensional indicator vectors to predict driver path choices and reconstruct missing trajectories. Validation with real-world LPR data from Wujin District demonstrates the algorithm's superior composite accuracy of 75.0 %, significantly outperforming the shortest path method (42.1 %) and an autoencoder model (38.0 %), confirming its accuracy and generalization capability.