HCHaechan Cho

Haechan Cho

*This is an algorithmically generated author record based on publicly available information.

Haechan Cho has a total of 4 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Dual-task urban traffic state estimation in camera undetected area using Swin Transformer and gated convolution

    by Haechan Cho, Hwasoo Yeo

    Camera sensors are increasingly deployed across urban road networks for traffic monitoring, yet their limited spatial coverage restricts their effectiveness. Estimating traffic states between cameras on urban arterials is therefore essential for advanced traffic control and vehicle routing. However, existing data-driven deep learning approaches have largely been designed in highway settings. Moreover, these methods predominantly rely on convolutional neural networks, which are limited in capturing the long-range spatiotemporal dependencies necessary for modeling complex urban traffic dynamics. To address these limitations, we propose DT-SwinTSE, a dual-task network that jointly estimates traffic speed and queue profile maps within a unified architecture. The proposed dual-task convolutional Swin Transformer block explicitly models cross-task feature interactions, enabling queuing dynamics to be embedded into the traffic state estimation (TSE) process. The network integrates Swin Transformer layers to capture long-range dependencies without border artifacts and gated convolutional layers to dynamically filter invalid pixels within camera-undetected regions. DT-SwinTSE is trained exclusively on simulated traffic data generated under diverse traffic demand and signal control conditions, allowing the model to learn essential urban traffic behaviors while mitigating the scarcity of labeled real-world data. Simulation-to-real transferability is evaluated using the NGSIM Peachtree and Lankershim benchmarks. Experimental results show that DT-SwinTSE consistently outperforms kernel-based imputation, data assimilation, and data-driven deep learning baselines—including the previous state-of-the-art U-Net—and maintains its superior performance under single-camera configurations and varying camera coverage, demonstrating strong robustness and adaptability. Simulation-based evaluations further demonstrate that DT-SwinTSE generalizes effectively across diverse traffic demand, signal timing, and simulation parameter configurations.