Developing a Self-Guided Field Robot for Greenhouse Asparagus Monitoring

Abstract

<b><sc>Abstract.</sc></b> Asparagus (Asparagus officinalis L.) is an important economic crop in Taiwan. Under a subtropical climate, the mother stalk method enhances the photosynthesis of plants to increase their growth and sustains a nearly year-round production. This method requires a lot of manpower to control the number of mother stalks during the harvest period because too many stalks will compete for the nutrients of the spears and lower their production in the field. However, the agricultural labor shortage has become a severe problem for farm management. An automatic monitoring system would be a solution to reduce the manpower for field management and provide the information, including the growth status of spears, the potential yield, and the best harvest time. The previous studies in our group have built a feasible deep learning model for asparagus growth identification. This study will focus on developing a navigation system for the field robot. The robot consisted of four-wheel-drive motors, an embedded controller, and a webcam. The navigation system was built by a deep convolutional neural network (DCNN) and image processing methods. A total of 729 front-view images were acquired and processed with semantic annotation for two classes, crop rows, and the lane. The ENet structure was selected as the real-time semantic segmentation model to identify the subject classes in the scene. Then, the Hough transform was introduced to obtain the left and right borderlines of the lane, and a guideline could be calculated as the suggested direction for driving adjustment. The developed ENet model achieved the mean intersection over union (mIoU) of 88.14% for the lane, and the mean absolute error (MAE) of the angle bias was 7.09 degrees at around 9.5 fps. The preliminary result showed the feasibility of using the DCNN to establish a self-guided field robot for greenhouse asparagus monitoring.