Visão computacional aplicada a automação de colhedora multifuncional de hortícolas - alface

Abstract

Agricultural automation has become significant in our country due to the need for domestic companies to compete properly with foreign companies and increased productivity and reduced losses. Using advanced technology features such as embedded electronic devices and precision farming techniques and process control, increased production is now possible. Horticultural crops are of great economic, social and food importance to the world's population. The current major challenge in the horticultural agricultural mechanization sector is harvesting. To this end, the development of a computer vision system employing image processing and artificial neural networks that allow the identification of vegetables and provide positioning parameters for a robotic actuator set for harvesting operations would meet this challenge. The objective of this work was to develop a detection system that through these resources enables the detection of vegetables and their positioning. For this purpose a capture apparatus was defined using model OV2640 serial cameras with ArduCam / Arduino control interface set in stereo vision configuration. The programming of the system was done in C # using the EMGU library using deep learning algorithms (YOLO) and depth metric algorithm by means of distance estimation map. As a result we have the definition of the apparatus where operating conditions were established and the performance of the image capture algorithms were analyzed, an image bank was collected, an artificial neural network was defined and training was performed with these images to identify the lettuce horticulture. , the object detection system was generated, the distance detection module was calibrated and lettuce detection tests were performed. Evaluating the results we concluded that it is possible to detect vegetables using the implemented system.