Detection of the centerline of rice seedling belts based on region growth sequential clustering-RANSAC

Automatic navigation with machine vision can improve the intelligence of agricultural robots in farmland. Accurate and rapid detection of crop rows can greatly contribute to the extraction of navigation lines during visual navigation. However, the detection of rice seedling centerline can be susceptible to the image noise caused by light changes, and machine vibration. In this study, a new approach was proposed to detect the centerlines of the rice seedling belt using regional growth order clustering-RANSAC (Random Sample Consensus, RANSAC). The rice seedlings were set at the regreening and early tillering stage. Firstly, the rice seedling images were acquired by the camera on the field patrol robot, and then divided into the paddy field background and seedling region using the normalized super green feature method and the maximum variance between classes method. The noise points in the seedling images were removed using the morphological method of the first etching and then opening operation. Secondly, the image was divided into 20 horizontal strips. The centroids of the rice pixel regions in the horizontal strips were taken as the feature points of rice seedlings, in order to reduce the amount of calculation for the high running speed. Thirdly, the feature points of the three horizontal strips at the bottom of the image were selected as the initial seeds, whereas, the feature points of rice seedlings were clustered by the regional growth sequence clustering method. The key parameters of the growth criteria were obtained to distinguish the crop rows using vertical projection accumulation in the binary image pixels of each horizontal strip. A series of experiments were carried out with numerous images of rice seedlings. The thresholds of expansion and distance were determined in the two critical periods of regreening and early tillering. As such, the seedling feature points of the same rice row were accurately grouped into the same category, according to the growth criteria. Finally, the centerlines of the seedling belts were fitted to obtain the visual navigation baselines of the field patrol robot using the RANSAC algorithm. The images of rice seedlings were obtained at the returning green and the early tillering stage using static and dynamic acquisition, in order to verify the real-time performance and adaptability of this model. The images included four lighting conditions: sunny, cloudy, front lighting, and backlighting. The results showed that this model performed better under different lighting conditions for both acquisitions. In addition, 400 images were randomly selected from two growth stages for comparison. Among them, the images of each growth stage included 100 on sunny days and 100 on cloudy days. The detection rate was above 97% for the centerlines of rice seedlings, while the average error angle was less than 2.34°, and the average detection time of each frame image was less than 15.53 ms. There was a great detection speed and accuracy in the rice seedling centerline extraction, compared with the YOLOv3 target detection. The seedling row was also extracted using regional growth mean and shift clustering. Generally, high adaptability and real-time performance were achieved in the rice growth stages, lighting conditions, and image acquisitions. The finding can fully meet the requirements of low-cost, lightweight computing, and high real-time performance for the navigation of field patrol robots.