Thermal imaging, when coupled with deep-learning-based analysis, can significantly improve highway work zone management and safety, though data scarcity in this field has been a historical challenge. We collected over 440 hours of thermal footage from highway sites in Massachusetts and New Hampshire and created a vehicle segmentation dataset featuring over 14,000 vehicle instances. We implement effective training strategies for deep learning models and develop algorithms to analyze vehicle trajectories and merging behaviors. Various visualization techniques are illustrated to better understand the extensive analytics data generated for each site, which can lead to development of data-driven strategies to improve Traffic management and safety. GitHub repository: https://github.com/z00bean/SmartWorkZoneControl.