Analysis Of Deep Learning-Based Visual Perception Technology for Picking Robots

Abstract

With global population growth and rising labor costs, the development of agricultural picking robots is crucial to improving agricultural productivity and reducing costs. A combination of deep learning technology and traditional vision technology has revolutionized the visual perception capability of agricultural picking robots. This has enabled robots to achieve more accurate fruit identification and localization in complex environments. In this paper, we review the visual perception technology of picking robots based on deep learning, which is primarily divided into two core technologies for fruit recognition and 3D reconstruction and localization during fruit picking. Firstly, we analyze the application of deep learning models in fruit recognition. We discuss how to integrate and process the feature information of fruits through deep learning neural networks to improve recognition accuracy and review the development of more typical deep learning models in the past three years. Secondly, we discuss the advantages and disadvantages of several types of traditional stereo vision technology. We synthesize the advantages and disadvantages of several types of current stereo vision technology that are more widely used. We also analyze how deep learning technology has optimized the model and combined it with traditional stereo vision technology in recent years to achieve the three-dimensional reconstruction and precise positioning of fruits. Finally, we summarize the main challenges of current picking robots, mainly the accuracy of the deep learning network, the difficulty of acquiring and calibrating the original image set, and the generality of the models carried by the picking robots. Additionally, we look forward to the future development direction.