Aiming at the complexity and diversity of the characteristics of field crop rows, the lack of robustness of the traditional crop row detection method, and the difficulty of parameter adjustment, a field crop row detection method based on feature engineering was proposed. Taking the seedling cotton crop row canopy as the recognition object, the crop row canopy characteristics were analyzed, and the feature expression model of the canopy of cotton crop was established with RGB image and depth image as the data source. The key feature parameters of crop row canopy were extracted by using feature dimensionality reduction method to reduce the amount of computation. A crop canopy feature segmentation model was established based on support vector machine technology to extract crop feature points. The method of crop row centerline detection was established by combining random sample consensus algorithm and principal component analysis. Using cotton crop row images with different illumination, weed and camera positions as test data, SVM classifiers with linear, RBF, and polynomial kernels were employed to conduct crop row canopy segmentation experiments. The performance of typical Hough transform, linear square method and the established crop row centerline detection method was compared and analyzed. The results showed that the RBF classifier had the best segmentation accuracy and robustness. The accuracy and speed of the established crop row centerline detection method were the best. The mean value of heading angle deviation was 0.80° and the standard deviation was 0.73°;the mean value of lateral position deviation was 0.90 pixels and the standard deviation was 0.76 pixels;the mean value of centerline fitting time was 55.74ms/f and the standard deviation was 4.31ms/f. The research results can improve the adaptability of crop row detection model, reduce the workload of parameter adjustment, and provide accurate navigation parameters for navigation system.