This study presents a method for size estimation of potato tubers under different clustering conditions. Images of potato tubers on a conveyor at lab in Dalhousie University, Department of Engineering, were captured using a commercially available RGB camera mounted above the conveyor. The captured images were labelled and fed to a Mask R-CNN (Region-based Convolutional Neural Network) deep learning architecture, which enabled instance segmentation of potato tubers even in cases of severe clustering (i.e., more than 100 tubers per frame) and mutually occluding tubers. Given that not all tubers were fully visible, image processing techniques were employed to randomly choose five fully visible potato tubers per frame. The minor and major diameters of the chosen tubers were calculated by fitting an ellipse over the detected potato mask and outputting the lengths of the minor and major axes.

Our method was demonstrated by using six images for each of the following clustering conditions: dense (100 - 200 tubers per frame), moderate (50 - 100 tubers per frame), and sparse (1 - 50 tubers per frame). The coefficient of determination for the major and minor diameters was 0.94 and 0.78 for all clustering scenarios, respectively, with the dense scenario having coefficients of 0.89 and 0.76. This technique allows for measuring minor and major diameters under varying clustering conditions without manual sampling and could replace the traditional size grading methods.

The full autonomy of agricultural machines will require human supervisors to monitor these machines during operation with the purpose of minimizing system failures or malfunctions. It is important, however, for a supervisor to recognize the emergency immediately in order to intervene. Warning devices typically rely on humans to perceive visual, auditory, and tactile signals. These warning methods, however, differ in their ability to elicit a rapid response. As a result, it is crucial to determine which warning method will be most effective in drawing the attention of a supervisor whenever an autonomous machine is in a state of emergency or danger. To achieve this objective, participants were recruited and asked to interact with a simulation of an autonomous sprayer. Two bimodal warning methods (audio-visual and visual-tactile sensory cues) in in-field and close-to-field remote supervision scenarios were considered in this study. Effectiveness of the bimodal warnings will be assessed based on the level of noticeability (which is a measure of level one situation awareness) and response time. The experimental study is currently in progress; we hope to be able to share preliminary findings during the conference. This study will help improve the performance of remote supervisors and minimize unexpected incidents during field operations with autonomous agricultural machines.