RESUMEN

Vine disease detection is considered one of the most crucial components in precision viticulture. It serves as an input for several further modules, including mapping, automatic treatment, and spraying devices. In the last few years, several approaches have been proposed for detecting vine disease based on indoor laboratory conditions or large-scale satellite images integrated with machine learning tools. However, these methods have several limitations, including laboratory-specific conditions or limited visibility into plant-related diseases. To overcome these limitations, this work proposes a low-altitude drone flight approach through which a comprehensive dataset about various vine diseases from a large-scale European dataset is generated. The dataset contains typical diseases such as downy mildew or black rot affecting the large variety of grapes including Muscat of Hamburg, Alphonse Lavallée, Grasa de Cotnari, Rkatsiteli, Napoca, Pinot blanc, Pinot gris, Chambourcin, Feteasca regala, Sauvignon blanc, Muscat Ottonel, Merlot, and Seyve-Villard 18402. The dataset contains 10,000 images and more than 100,000 annotated leaves, verified by viticulture specialists. Grape bunches are also annotated for yield estimation. Further, tests were made against state-of-the-art detection methods on this dataset, focusing also on viable solutions on embedded devices, including Android-based phones or Nvidia Jetson boards with GPU. The datasets, as well as the customized embedded models, are available on the project webpage.

RESUMEN

In this paper, an efficient normal estimation and filtering method for depth images acquired by Time-of-Flight (ToF) cameras is proposed. The method is based on a common feature pyramid networks (FPN) architecture. The normal estimation method is called ToFNest, and the filtering method ToFClean. Both of these low-level 3D point cloud processing methods start from the 2D depth images, projecting the measured data into the 3D space and computing a task-specific loss function. Despite the simplicity, the methods prove to be efficient in terms of robustness and runtime. In order to validate the methods, extensive evaluations on public and custom datasets were performed. Compared with the state-of-the-art methods, the ToFNest and ToFClean algorithms are faster by an order of magnitude without losing precision on public datasets.

Asunto(s)

RESUMEN

A novel method is proposed for the absolute pose estimation of a central 2D camera with respect to 3D depth data without the use of any dedicated calibration pattern or explicit point correspondences. The proposed method has no specific assumption about the data source: plain depth information is expected from the 3D sensing device and a central camera is used to capture the 2D images. Both the perspective and omnidirectional central cameras are handled within a single generic camera model. Pose estimation is formulated as a 2D-3D nonlinear shape registration task which is solved without point correspondences or complex similarity metrics. It relies on a set of corresponding planar regions, and the pose parameters are obtained by solving an overdetermined system of nonlinear equations. The efficiency and robustness of the proposed method were confirmed on both large scale synthetic data and on real data acquired from various types of sensors.

RESUMEN

We consider a robot that must sort objects transported by a conveyor belt into different classes. Multiple observations must be performed before taking a decision on the class of each object, because the imperfect sensing sometimes detects the incorrect object class. The objective is to sort the sequence of objects in a minimal number of observation and decision steps. We describe this task in the framework of partially observable Markov decision processes, and we propose a reward function that explicitly takes into account the information gain of the viewpoint selection actions applied. The DESPOT algorithm is applied to solve the problem, automatically obtaining a sequence of observation viewpoints and class decision actions. Observations are made either only for the object on the first position of the conveyor belt or for multiple adjacent positions at once. The performance of the single- and multiple-position variants is compared, and the impact of including the information gain is analyzed. Real-life experiments with a Baxter robot and an industrial conveyor belt are provided.