In response to the problems in traditional manual counting methods such as low efficiency and high risk, a steel plate counting method based on peak-valley features and combination voting was designed. Frame difference method was used to detect the region of interest and perform image preprocessing and edge detection on it. Then, the edge image of each cut steel plate was used for vertical projection to obtain the gray projection curve and remove noise. The first-order forward difference and second-order forward difference were used to obtain the peak and valley points of the steel plate from the projection curve, and then the peak and valley points of multiple images were combined by voting to obtain the number of steel plates. Experimental results show that compared with manual counting and other point expansion algorithms, the accuracy of this method is higher than 95%, which can meet actual business needs.