Roadside-based multiple object tracking(MOT) is a crucial and fundamental research field in computer vision, providing abundant information for autonomous driving and intelligent transportation systems. However, in real-world tracking applications, frequent movements of traffic participants and variations in lighting conditions caused by different weather conditions pose significant challenges for roadside tracking. To overcome these challenges, we propose an online one-shot framework for roadside multiple object tracking, called ROMOT. we firstly propose the DLA-SA network in order to efficiently extract vehicle features. In addition, we introduce a fusion matrix called EC matrix and leverage low score detection information through a cascade matching method to address the correlation problem of long-term lost trajectories. We conduct experiments on the UA-DETRAC dataset to evaluate the performance of ROMOT, where ROMOT achieves a remarkable MOTA of 71.0% for the tracked objects, while maintaining an inference speed of 25Hz, meeting the real-time requirements.