Aiming at the path planning in rice harvesting visual navigation, a method for extracting the boundary line of rice harvesting area was presented. Camera distortions were eliminated by camera calibration. According to the ultra-red characteristics model 2R-G-B, binary images were carried out based on integrated threshold method. Noise was eliminated by using the opening operation and closing operation in morphology. The ROI area was dynamically set according to the image grayscale vertical projection value. Crop line fitting key points were obtained by horizontal scanning. The boundary line of rice to be harvested was extracted by multi-segment cubic B-spline curve fitting method. Laboratory tests showed that the average error of distance information extracted based on the proposed image processing method was 9.9mm, the deviation rate was 2.0%, the average error of angle information was 0.77°, and the error rate was 2.7%. The field experiment of harvesting path extraction was carried out for two crops, Zhongjing 798 and Lindao 20 under four different light environments，direct sunlight, backlight, strong light and weak light，respectively. For the harvest image of Zhongjing 798, the average pixel error of crop line key point recognition was 28.7 pixel, the average distance error was 39.7mm, and the average relative error was 2.7%. The minimum average pixel error was 26.2 pixel in strong light, the minimum average distance error was 23.9mm in weak light, the minimum average relative error was 2.0% in strong light, in direct sunlight the algorithm was most stable with a standard deviation of 6.8 pixel. For the harvested image of Lindao 20, the average pixel error of crop line key point recognition was 36.5 pixel, the average distance error was 45.0mm, and the average relative error was 2.8%. All minimum value of each indicator was in backlight, the minimum average pixel error was 29.5 pixel, the minimum average distance error was 36.9mm, the minimum average relative error was 2.3% and the minimum standard deviation was 10.8 pixel. The average processing time of single frame image was 38ms, which can provide reference for crop line detection and automated navigation in harvesting.