Search

Traffic Sign Recognition (TSR) is one of the many utilities made possible by embedded systems with internet connections. Through the usage of vehicular cameras, it's possible to capture and classify traffic signs in real time with Artificial Intelligence (AI), more specifically, Convolutional Neural Networks (CNNs) based techniques. This article discusses the implementation of such TSR systems, and the building process of datasets for AI training. Such datasets include a brand new class to be used in TSR, vegetation occlusion. The results show that this approach is useful in making traffic sign maintenance faster since this application turns vehicles into moving sensors in that context. Leaning on the proposed technique, identified irregularities in traffic signs can be reported to a responsible body so they will eventually be fixed, contributing to a safer traffic environment. This paper also discusses the usage and performance of different YOLO models according to our case studies.

• 90 videos were recorded during harvest at the harvester conveyor belt. • Different input resolutions were assessed for the YOLOv3-tiny detection system. • The georeferenced video frames were used for object detection. • Coffee maturation stage was mapped during harvest. In this study, an algorithm is implemented with a computer vision model to detect and classify coffee fruits and map the fruits maturation stage during harvest. The main contribution of this study is with respect to the assignment of geographic coordinates to each frame, which enables the mapping of detection summaries across coffee rows. The model used to detect and classify coffee fruits was implemented using the Darknet, an open source framework for neural networks written in C. The coffee fruits detection and classification were performed using the object detection system named YOLOv3-tiny. For this study, 90 videos were recorded at the end of the discharge conveyor of a coffee harvester during the 2020 harvest of arabica coffee (Catuaí 144) at a commercial area in the region of Patos de Minas, in the state of Minas Gerais, Brazil. The model performance peaked around the ~3300th iteration when considering an image input resolution of 800 × 800 pixels. The model presented an mAP of 84%, F1-Score of 82%, precision of 83%, and recall of 82% for the validation set. The average precision for the classes of unripe, ripe, and overripe coffee fruits was 86%, 85%, and 80%, respectively. As the algorithm enabled the detection and classification in videos collected during the harvest, it was possible to map the qualitative attributes regarding the coffee maturation stage along the crop lines. These attribute maps provide managers important spatial information for the application of precision agriculture techniques in crop management. Additionally, this study should incentive future research to customize the deep learning model for certain tasks in agriculture and precision agriculture. [ABSTRACT FROM AUTHOR]