Evaluation of a Stereo Vision System for Cotton Row Detection and Boll Location Estimation in Direct Sunlight

Abstract

Cotton harvesting is performed by using expensive combine harvesters which makes it difficult for small to medium-size cotton farmers to grow cotton economically. Advances in robotics have provided an opportunity to harvest cotton using small and robust autonomous rovers that can be deployed in the field as a “swarm” of harvesters, with each harvester responsible for a small hectarage. However, rovers need high-performance navigation to obtain the necessary precision for harvesting. Current precision harvesting systems depend heavily on Real-Time Kinematic Global Navigation Satellite System (RTK-GNSS) to navigate rows of crops. However, GNSS cannot be the only method used to navigate the farm because for robots to work as a coordinated multiagent unit on the same farm because they also require visual systems to navigate, avoid collisions, and to accommodate plant growth and canopy changes. Hence, the optical system remains to be a complementary method for increasing the efficiency of the GNSS. In this study, visual detection of cotton rows and bolls was developed, demonstrated, and evaluated. A pixel-based algorithm was used to calculate and determine the upper and lower part of the canopy of the cotton rows by assuming the normal distribution of the high and low depth pixels. The left and right rows were detected by using perspective transformation and pixel-based sliding window algorithms. Then, the system determined the Bayesian score of the detection and calculated the center of the rows for the smooth navigation of the rover. This visual system achieved an accuracy of 92.3% and an F1 score of 0.951 for the detection of cotton rows. Furthermore, the same stereo vision system was used to detect the location of the cotton bolls. A comparison of the cotton bolls’ distances above the ground to the manual measurements showed that the system achieved an average R2 value of 99% with a root mean square error (RMSE) of 9 mm when stationary and 95% with an RMSE of 34 mm when moving at approximately 0.64 km/h. The rover might have needed to stop several times to improve its detection accuracy or move more slowly. Therefore, the accuracy obtained in row detection and boll location estimation is favorable for use in a cotton harvesting robotic system. Future research should involve testing of the models in a large farm with undefoliated plants.