Combinando métodos de detecção de objetos com sistema de localização e mapeamento simultâneos

Abstract

This work explored the combination of Deep Learning (DL) and Simultaneous Localization and Mapping (SLAM) to enhance precision agriculture, with a focus on detecting and estimating the distance to apples in orchards. A thorough literature review was conducted, analyzing approaches that integrate deep neural networks with traditional SLAM methods, identifying promising applications in various fields, including agriculture. Data collection was performed using a stereoscopic camera, capturing images with depth information. Bounding boxes were manually annotated around visible apples, and the MinneApple dataset was added to enhance model generalization. We trained 12 variations of the YOLO architecture, with YOLOv5x achieving the best performance, reaching 0.861 in F1-Score on the validation set. An algorithm was developed to estimate the distance to each detected apple, integrating it into the YOLO detection pipeline. The results demonstrated the accuracy and real-time viability of the system, allowing efficient detection and distance estimation of apples. This work contributes to the evolution of the combination of DL and SLAM, opening new research prospects for automation and robotics, particularly in fruit orchard monitoring and harvesting applications.