Aiming at the problems that orchard roads have no obvious boundaries and there are shadows, soil and sand interference at the edges of the road, a recognition method of orchard unstructured roads based on feature fusion was proposed. The distortion parameters were obtained through camera calibration to correct the distortion of the acquired image, and a dynamic region of interest (ROI) extraction method based on the combination of filtering and gradient statistics was proposed to select the ROI of the S component of the HSV color space. The maximum value method was used to merge the color features with the S component mask for multidirectional texture features for binarization and noise reduction. The feature points were found according to the abrupt features of road edges, and a two-level pseudo feature points elimination method based on the dual constraints of distance and position was proposed. To better fit the irregular edges of unstructured road, the method of segmentation cubic spline interpolation was introduced to fit the road edges to realize road recognition. The experimental results showed that under the six working conditions of sunny day, cloudy day, straight light, backlight, sunny day in winter and rain and snow weather, the mean value of average longitudinal deviations of S component, texture image and fusion image were 2.43 pixels, 39.71 pixels and 1.36 pixels, respectively, and the mean value of average deviation rates were 0.99%, 18.02% and 0.54%, respectively. Compared with the S component and texture image, the average longitudinal deviation and average deviation rate of the fusion image constructed by this method were effectively reduced. The mean value of average deviations of least squares method, random sample consensus method (RANSAC) and segmentation cubic spline interpolation method for fitting edges were 2.64 pixels, 3.16 pixels and 0.66 pixels, respectively, the mean value of average deviation rates were 1.02%, 1.21% and 0.26%, respectively, and the average standard deviations of deviation rate were 0.23%, 0.31% and 0.09%, respectively. The mean value of average deviation, mean value of average deviation rate and average standard deviation of deviation rate of the algorithm were the minimum, which indicated that the fitting method had higher fitting accuracy and better stability. Under the six working conditions, the average processing time of a single image of this algorithm was 89.9 ms, which met the real-time requirements of agricultural robots in the process of operation. The method can provide a reference for agricultural robots to recognize unstructured roads in complex orchard environments.