Junjie Hu (胡俊杰)

Designed a high-order traffic state reconstruction method demonstrating traffic flow's Markov property. Leveraged Markov matrix prior knowledge for high-accuracy traffic imputation tasks using an auto-encoder framework.

Created a crash data enhancement strategy utilizing historic priori spatial and temporal knowledge. Developed a novel method, Priori Spatial and Temporal Causal Graph Convolutional Network (PST-CGCN), for predicting crash risk based on causal inference and graph convolutional networks.

Designed a new data-driven and transferable analysis model, the Tucker-Net based Susceptible-Infectious-Recovered-Susceptible (SIRS) model (TNBSM), for exploring dynamic correlations within crash zones using tensor decomposition and the SIRS model.

Applied a coarse-grained phase space algorithm, introduced complex network technology into vehicle-following behavior analysis, and revealed distinct characteristics and differences between autonomous vehicles (AV) and human-driven vehicles (HV).

Introduced an adaptive trajectory-based visibility graph (TVG) framework, a novel method for dissecting vehicle dynamics by transforming planar trajectory data into complex networks. This framework features a tunable visibility tolerance coefficient, dynamically scaled by lateral displacement, enabling the TVG to capture geometric occlusions and maneuver-specific spatial scales.<

Proposed a modified K-Nearest Neighbors (KNN) framework for reconstructing and imputing Time-Space Diagrams (TSD) from sparse data by matching similar spatiotemporal neighborhood features defined by Queen Contiguity Spatial Rule.

Revisited an end-to-end approach for driving style recognition by exploring driver behavior heterogeneity through trajectory prediction model errors. Utilized spatial attention and convolutional social pooling to learn interdependencies in vehicle motion and introduced a multi-modal distribution for future trajectories based on driving style.

Posited that residuals of car-following (cf) model are a composite of structural errors and random error and utilized an Unobserved Components Model to decompose the residual series from three calibrated CF models.

Introduced a plug-and-play framework named Residual-Aware Meta-learning Re-weighting (RAMR), which reframes the Car-following model training problem as a bilevel optimization task. Comparative experiments were conducted on four datasets, applying three classical physical CF models and three data-driven predictors to comprehensively validate the efficacy and versatility of RAMR.

Using a global dataset of 70 cities, Contributed to investigating how road network morphology, quantified by directional entropy, affects key metrics of urban sustainability and resilience, including CO₂ emissions, commute times, and congestion.