| description abstract | Matching trajectory segments from different roadside sensors to form continuous, regional-level trajectories has become an emerging topic for cooperative vehicle infrastructure systems (CVIS). The current appearance-based matching method faces both subtle interinstance discrepancies and significant intrainstance differences in vehicle images. This paper proposes a spatiotemporal and appearance-based cross-domain trajectory-matching (STA-CDTM) framework for distributed cameras in the urban road network, which comprises two modules: spatiotemporal feature calculation and fusion feature matching. The spatiotemporal feature calculation module employs a dynamic graph-based spatiotemporal propagation model (DG-STPM) to model the traffic propagation characteristics within the road network and extract the spatiotemporal correlation features of the vehicle trajectory. The fusion feature matching module utilizes a spatiotemporal mask (STM) to determine the candidate nodes and trajectories through the spatiotemporal constraints and employs the enhanced similarity metric (ESM) for trajectory matching by fusing the spatiotemporal correlation and appearance features. On the public pNEUMA data set, the proposed DG-STPM demonstrated reductions in mean absolute error (MAE) by 15.4%, 10.8%, and 5.1% in free flow, congestion, and jam scenarios, respectively, effectively capturing traffic propagation characteristics. Additionally, the framework achieved the highest identification F-score (IDF1) and identification precision (IDP) scores of 0.8251, and 0.8567 in trajectory matching, tested on a field-collected roadside camera data set from Jiading, Shanghai, China, indicating superior performance across various traffic conditions. | |
