High-resolution reconstruction of non-motorized trajectory in shared space: A new approach integrating the social force model and particle filtering

Abstract

High-resolution trajectory data from the intersection shared space (ISS) is an ideal resource for testing autonomous driving and studying microscopic traffic flow theory in urban road traffic environments. It is common to use computer vision (CV) technologies to extract trajectory data from camera recordings at the ISS. However, it is challenging to accurately extract non-motorized (NONM) trajectories due to vehicle occlusion and mutual interference, which cause trajectory noise and missing. This study proposed a novel NONM trajectory reconstruction method, which integrates the social force model with the particle filter to reduce trajectory noise. The approach was tested on trajectories extracted from two advanced CV algorithms (ie, Yolov7 and MASK RCNN), and its performance was compared with seven SOTA methods. The results show that the proposed method achieved a much better RMSE (0.6 meters) than both the baseline (unprocessed CV trajectories, the RMSE is 1.16 meters) and selected SOTA methods (the best RMSE is 0.78 meters). The contribution of the study is to enhance the classical noise reduction algorithm by introducing the driving interaction model; ablation experiments and sensitivity analysis further demonstrate the importance of the driving interaction model in reducing trajectory noise and the stability of the method. The approach also includes an LSTM-Hungarian part to make up the missing trajectory. The proposed method is expected to serve as a useful post-processing tool for current CV algorithms.