An approach of robust feature recognition and precise path tracking based on two visual field windows was proposed for an AGV to identify multibranch paths and station point reliably, and to follow guide paths accurately. The whole visual field was used as a pattern recognition window, in which a recognition method based on kernel principal component analysis (KPCA) and BP neural network was developed. Path features were mapped to a highdimensional space by using the kernel function and then their dimensionalities were reduced by using PCA. After dimensionality reduction, the sample matrices were recognized by utilizing a BP neural network. Besides, a scaling window method based on a vertical view angle and a tilt installation angle of a camera was suggested for a guidance scanning window. In this window, guide paths were simplified according to a linear model and fitted by using the least square method. Path deviations with respect to the fitted straight line were estimated for AGV guidance. Experimental results show that the KPCA-BP approach improves the realtime performance and robustness of path feature recognition significantly, the average correct rate of which is 99.5% for six types of landmark feature, and that the guidance scanning window decreases the computing error resulted from linear fitting of guide paths effectively, the tracking error of which is no more than 3mm for linear path and 30mm for curvilinear path.