Accurate metro passenger detection is considered a fundamental technique for the smart metro station. The major challenge is false positives (FPs) when hunting for passengers in metro carriages and stations. This article provides an in-depth analysis to reveal the reasons for this problem in the two-stage convolutional neural network (CNN)-based detector and proposes a novel two-mode refined proposals algorithm to address the problem, which includes cascade mode and intrinsic information mode, or CRP and IRP algorithms for short. Aided by the two-mode refined Proposals algorithm, this article designs two small but fast and accurate pedestrian detectors: MetroNext-CRP and MetroNext-IRP to meet the application requirements of different tasks. Based on various challenging benchmark datasets and two metro scene datasets, the experimental results have demonstrated that the two-mode refined proposals algorithm is effective and can improve pedestrian detection accuracy by removing the FPs. Compared with the existing state-of-the-art detection networks, MetroNext-CRP achieves competitive detection results with acceptable computational cost and MetroNext-IRP demonstrates better detection accuracy with fast inference speed and without extra memory consumption, thus providing a practical solution for pedestrian detection tasks on various hardware platforms, particularly tailored to edge devices. Finally, the two-mode MetroNexts are deployed in a metro’s onboard computer, and the experiments in real-life campus and metro scenes further demonstrate the feasibility of the two detectors, especially the field experiments in metro scenes strongly support that they can be served as a smart instrument to provide accurate passenger position information to metro operators in complicated and realistic metro scenes.