Efficiently recognizing kilometer markers in railway systems is crucial for ensuring the safety and reliability of train operations, particularly within the Industrial Internet of Things (IIoT) framework where edge devices are often resource-constrained. This paper highlights the significance of a real-time AI-driven approach to railway kilometer marker recognition. We introduce an innovative method that employs spatio-temporal slimming and multi-view fusion techniques, elevating both precision and computational efficiency for real-time analytics in IIoT. Our approach begins with the implementation of an Adaptive Region of Interest (AROI) and a spatio-temporal calibration mechanism for effective marker detection in real-time. Furthermore, a multi-view fusion method addresses challenges such as occlusion, blurriness, and low-light conditions, common in real-world industrial environments, which includes a variation-aware memory bank for constructing informative views and a fusion network. Experimental results demonstrate the effectiveness of our method in a real-world rail transportation setting, significantly enhancing the accuracy and efficiency of kilometer marker recognition, thereby contributing to the safety and operational efficiency of rail systems.