These years, using tractors and self-propelled agricultural machines has been caused to increase of human death numbers due to accident with them. So, having an alarm system on agricultural vehicles for humans and obstacles detection can be beneficial for reducing of accidents. This study aimed to develop a vision system for human detection to use in agricultural tractors. A camera was installed in front of the tractor, and an algorithm was designed in MATLAB for image processing. After trimming the input image, the human was identified using watershed and gradient magnitude techniques, and subtracting human from the background. According to results of the algorithm evaluation in various image acquisitions, the proposed algorithm had high accuracy at distances of ? to ?? meters, in comparison with two common human detection methods (Histogram of oriented gradient and Cascade). Average accuracy of the proposed algorithm in these distances was ?? to ?? percent. The results of analysis of variance for accuracy of human detection showed that the effect of identification method and distance on the accuracy of human detection were significant at the ?% level. Also the interaction effects between different parameters were significant.