To address the issue of continuous automatic measurement of elevator guide rail deformation, this paper proposes an online detection method for the straightness of elevator guide rails based on computer vision. The poor and dim environment at the top of the elevator interferes with the collection of rail video images. To address this, a Reflectance-Guided, Contrast-Accumulated Histogram Equalization (RG-CACHE) algorithm is used to enhance the rail images. Furthermore, in response to the high fusion of the elevator guide rail edges with the background and the low contrast, an edge extraction method based on the deep learning Segment Anything Model (SAM) is presented to accurately extract the guide rail edges. Finally, an improved ordinary least squares method (OLS) that can automatically eliminate obvious outliers is proposed for straightness detection, and the goodness of fit is used for straightness evaluation. Experimental results show that this method can accurately achieve online detection of the straightness of elevator guide rails.