Mechanical Wild Blueberry (Vaccinium angustifolium Ait.) Harvester Tote Volume Estimation Using Time-of-Flight Imagery for Automated Yield Monitoring

Abstract

This study investigates the development of a novel real-time volume estimation and yield monitoring system for deployment on a mechanical wild blueberry harvester. Specifically, the focus is to develop an automated solution for tote filling. The developed system can also produce yield maps, a crucial tool in agriculture that has not yet been implemented in the wild blueberry industry. A notable challenge this system addresses is the considerable variance in tote weight, which can exceed 30%. The developed yield monitoring solution integrates seamlessly with the industry's standard mechanized harvester, equipped with electro-hydraulic control. The volume estimation system utilizes a 3D time-of-flight (ToF) camera and an RGB colour camera. The two cameras were mounted directly to the tote filling conveyor using a set of linear rails and custom attachments. Further modifications were made to the mounting system through a custom linkage mounting apparatus to provide a further range of motion for the vision system, thereby facilitating a comprehensive visual assessment of the tote area. To estimate the volume of product, realtime segmentation of the tote itself from the harvested product was used through a neural network framework. This, in conjunction with a voxel grid surface reconstruction algorithm, effectively determined the volume of the tote contents with a Mean Absolute Error (MAE) of up to 10%. Through a neural network-based segmentation algorithm used in combination with a developed Keras 1-dimensional feedforward neural network, the changes in mass and debris measured within the tote were able to be estimated. Beyond the development of the system, two separate trials were run sequentially to assess the effectiveness of the volume estimation system and the yield monitoring system. These trials demonstrated the effectiveness of a ToF camera for volume estimation with an average error of 1.2%. Building on this foundational success, the culmination of these efforts produced a highly accurate (MAE of 3.4% of tote mass), real-time deployable automated tote filling and yield monitoring system, specifically engineered for the commercial wild blueberry harvester.