The work aims to study an automatic counting method of bundled round bars from the image of their bottom surface caught by a camera. The radius of each bar's bottom surface was extracted by the Hough transformation method for normalizing. Then the image was zoomed to make every circle have a standard radius. The sliding window method was used to determine whether the center of the detection window was the center of a round bar. The result was labeled on another image. SVM classifiers were used for detection. The counting result was obtained by counting the number of connected regions in the labeled image of marked centers. Finally, the effectiveness of the method was tested with images in the test library. The experiment results showed that the accuracy of the methods in this paper on the test library was 94%, which was much higher than ordinary template matching method. The accuracy of the method in this paper is improved greatly compared with ordinary method. It also has a good robustness, and can be applied to automatic counting system for bundled round bars.