Robotic data acquisition of sweet pepper images for research and development

Abstract

A main problem limiting the development of robotic harvesters is robust fruit detection [5]. Despite intensive research conducted in identifying the fruits and their location [2,3], current fruit detection algorithms have a limited detection rate of 0.87 which is unfeasible from an economic perspective [5]. The complexity of the fruit detection task is due to the unstructured and dynamic nature of both the objects and the environment [4-6]: the fruit have inherent high variability in size, shape, texture, and location; occlusion and variable illumination conditions significantly influence the detection performance[3]. A common practice for image processing R&D for complicated problems is the acquisition of a large database (e.g., Labelme open source labeling database [1], Oxford building dataset [2]). These datasets enable to advance vision algorithms development [7] and provide a benchmark for evaluating new algorithms. To the best of our knowledge, to date there is no open dataset available for R&D in image processing of agricultural objects. Evaluation of previously reported algorithms was based on limited data [5]. Previous research indicated the importance of evaluating algorithms for a wide range of sensory, crop, and environmental conditions [5]. A robotic acquisition system and procedure was developed using a 6 degree of freedom manipulator, equipped with 3 different sensors to automatically acquire images from several viewpoints with different sensors and illumination conditions. Measurements were conducted along the day and at night in a commercial greenhouse and resulted in a total of 1764 images from 14 viewpoints for each scene. Additionally, drawbacks and advantages of the proposed approach as compared to other approaches previously utilized will be discussed along with recommendations for future acquisitions.