Mejoramiento en tiempo real de mapas 3D poco densos mediante superpixeles

Claudia Cruz-Martinez (2016)

In this work we address the problem of 3D mapping of environments with large textureless regions, which generates sparse 3D maps that may not represent well the mapped scene. To deal with this problem, we propose to enhance sparse 3D maps by using a superpixel-based segmentation with the aim of generating denser 3D maps of the scene in real time. This can be exploited in virtual reality and robotics applications where rapid generation of 3D maps may be required. Superpixels are middle-level features, which represent a homogeneous regions in an image, that can be connected in order to segment untextured areas. Based on this, we propose a GPU architecture for: i) superpixel extraction considering chromatic information, ii) superpixel-based segmentation, generation of connectivity matrix to compute the connected components algorithm and iii) mapping of segmented regions to 3D points. We use the ORB-SLAM system to generate a sparse 3D map and to project the untextured segments onto it at 27 FPS. We assessed our approach in terms of the segmentation and map quality. Regarding the latter, covered area by the generated map, depth accuracy and computational performance are reported.

Este trabajo se enfoca en el problema del mapeo tridimensional de ambientes con grandes regiones sin textura, los cuales generan mapas 3D dispersos que pueden no representar de forma adecuada el ambiente en cuestión. Para lidiar con este problema, se propone el mejoramiento de mapas dispersos 3D mediante el uso de la segmentación basada en superpixeles con el objetivo de generar un mapa 3D más denso en tiempo real. Este tipo de sistemas pueden ser de mucha utilidad en aplicaciones de realidad aumentada y virtual además de la solución de problemas orientados a robótica; donde se requiere una generación rápida de mapas tridimensionales. Los superpixeles son características de nivel medio los cuales representan regiones homogéneas en una imagen, que pueden ser conectados a fin de segmentar aéreas sin texturas. Basándose en esto, se propone una arquitectura GPU para: i) la extracción de superpixeles utilizando información cromática y/o de profundidad, ii) segmentación basada en superpixeles mediante la generación de la matriz de conectividad para ejecutar el algoritmo de componentes conectados, y iii) mapeo de regiones segmentadas a puntos tridimensionales con coordenadas del mundo real. Se utiliza el sistema ORB-SLAM para generar un mapa 3D disperso y proyectar los segmentos sin textura sobre el mismo en 27 FPS. Se evalúa la propuesta en términos de calidad de la segmentación y del mapa generado. Con respecto a esto último, se reporta la cantidad de área cubierta por el mapa generado, la precisión de la profundidad estimada y el desempeño computacional.

Master thesis

Superpixel SLIC RGB-D images Segmentation SLAM 3D mapping GPU CIENCIAS FÍSICO MATEMÁTICAS Y CIENCIAS DE LA TIERRA MATEMÁTICAS CIENCIA DE LOS ORDENADORES SISTEMAS EN TIEMPO REAL

Color characterization comparison for machine vision-based fruit recognition

Jair Cervantes Canales Farid García Lamont ASDRUBAL LOPEZ CHAU JOSE SERGIO RUIZ CASTILLA (2015)

In this paper we present a comparison between three color characterizations methods applied for fruit recognition, two of them are selected from two related works and the third is the authors’ proposal; in the three works, color is represented in the RGB space. The related works characterize the colors considering their intensity data; but employing the intensity data of colors in the RGB space may lead to obtain imprecise models of colors, because, in this space, despite two colors with the same chromaticity if they have different intensities then they represent different colors. Hence, we introduce a method to characterize the color of objects by extracting the chromaticity of colors; so, the intensity of colors does not influence significantly the color extraction. The color characterizations of these two methods and our proposal are implemented and tested to extract the color features of different fruit classes. The color features are concatenated with the shape characteristics, obtained using Fourier descriptors, Hu moments and four basic geometric features, to form a feature vector. A feed-forward neural network is employed as classifier; the performance of each method is evaluated using an image database with 12 fruit classes.

Book part

Color characterization Fruit classification RGB images INGENIERÍA Y TECNOLOGÍA

Contrast enhacenment of RGB color images by histogram equalization of color vectors' intensities

Jair Cervantes Canales Farid García Lamont ASDRUBAL LOPEZ CHAU JOSE SERGIO RUIZ CASTILLA (2018)

Mejora del contraste de imagenes de color RGB

The histogram equalization (HE) is a technique developed for image contrast enhancement of grayscale images. For RGB (Red, Green, Blue) color images, the HE is usually applied in the color channels separately; due to correlation between the color channels, the chromaticity of colors is modified. In order to overcome this problem, the colors of the image are mapped to different color spaces where the chromaticity and the intensity of colors are decoupled; then, the HE is applied in the intensity channel. Mapping colors between different color spaces may involve a huge computational load, because the mathematical operations are not linear. In this paper we present a proposal for contrast enhancement of RGB color images, without mapping the colors to different color spaces, where the HE is applied to the intensities of the color vectors. We show that the images obtained with our proposal are very similar to the images processed in the HSV (Hue, Saturation, Value) and L*a*b* color spaces.

Article

Color characterization Histogram equalization RGB images INGENIERÍA Y TECNOLOGÍA

Evaluating maize genotype performance under low nitrogen conditions using RGB UAV phenotyping techniques

Maria Luisa Buchaillot Adrian Gracia-Romero Omar Vergara Diaz Mainassara Zaman-Allah Amsal Tarekegne Jill Cairns Prasanna Boddupalli Jose Luis Araus Shawn Kefauver (2019)

Maize is the most cultivated cereal in Africa in terms of land area and production, but low soil nitrogen availability often constrains yields. Developing new maize varieties with high and reliable yields using traditional crop breeding techniques in field conditions can be slow and costly. Remote sensing has become an important tool in the modernization of field-based high-throughput plant phenotyping (HTPP), providing faster gains towards the improvement of yield potential and adaptation to abiotic and biotic limiting conditions. We evaluated the performance of a set of remote sensing indices derived from red–green–blue (RGB) images along with field-based multispectral normalized difference vegetation index (NDVI) and leaf chlorophyll content (SPAD values) as phenotypic traits for assessing maize performance under managed low-nitrogen conditions. HTPP measurements were conducted from the ground and from an unmanned aerial vehicle (UAV). For the ground-level RGB indices, the strongest correlations to yield were observed with hue, greener green area (GGA), and a newly developed RGB HTPP index, NDLab (normalized difference Commission Internationale de I´Edairage (CIE)Lab index), while GGA and crop senescence index (CSI) correlated better with grain yield from the UAV. Regarding ground sensors, SPAD exhibited the closest correlation with grain yield, notably increasing in its correlation when measured in the vegetative stage. Additionally, we evaluated how different HTPP indices contributed to the explanation of yield in combination with agronomic data, such as anthesis silking interval (ASI), anthesis date (AD), and plant height (PH). Multivariate regression models, including RGB indices (R2 > 0.60), outperformed other models using only agronomic parameters or field sensors (R2 > 0.50), reinforcing RGB HTPP’s potential to improve yield assessments. Finally, we compared the low-N results to the same panel of 64 maize genotypes grown under optimal conditions, noting that only 11% of the total genotypes appeared in the highest yield producing quartile for both trials. Furthermore, we calculated the grain yield loss index (GYLI) for each genotype, which showed a large range of variability, suggesting that low-N performance is not necessarily exclusive of high productivity in optimal conditions.

Article

RGB UAV CIELab CIENCIAS AGROPECUARIAS Y BIOTECNOLOGÍA MAIZE NITROGEN PHENOTYPES AGRICULTURAL SCIENCES AND BIOTECHNOLOGY MAIZE NITROGEN PHENOTYPING REMOTE SENSING

Deficiencias de hierro y manganeso en hojas de frijol (Phaseolus vulgaris L) identificadas mendiante análisis textural, color de imágenes digitales y redes neuronales artificiales

EDGAR GARCIA CRUZ (2013)

Tesis (Maestría en Ciencias, especialista en Edafología).- Colegio de Postgraduados, 2013.

En la presente investigación se analizaron imágenes digitales de hojas de frijol (Phaseolus vulgaris L.) para identificar con un clasificador, deficiencias de hierro (Fe) y manganeso (Mn). A los 24 días después de la siembra (dds) se les suministró la solución nutritiva de acuerdo a ocho tratamientos: dos deficiencias parciales, una de 50 % Fe y otra de 50 % Mn; dos deficiencias totales totales, 0 % Fe y una más de 0 % Mn además de una interacción (0 % Fe, 0 % Mn) y dos dosis excedentes (200 % Fe y 200 % Mn); finalmente un tratamiento testigo (100 % Fe, 100 % Mn) usando como referencia la solución Steiner. A partir de imágenes digitales de muestras de hojas de los tratamientos obtenidas a los 63 dds, se calcularon variables de color con los valores promedio de los canales de los espacios de color RGB y CIELab, el croma y el matiz. Además, se calcularon promedios de cuatro variables texturales: Segundo momento angular (SMA), entropía (EN), inercia (IN) y homogeneidad local (HoL). Tanto las variables de color como las de textura fueron usadas como variables independientes para generar clasificadores de grupos de datos de ocho, seis y cuatro tratamientos o clases de salida de deficiencias de Fe y Mn con el programa de redes neuronales NeuroShell® Classifier Release 2.2. Se obtuvo que usando sólo características texturales se obtienen bajos porcentajes de precisión en los clasificadores. Usando sólo características de color mejoran los porcentajes de clasificaciones correctas pero se requiere un mayor uso de neuronas de la capa interna, mientras que la combinación de características texturales y de color genera mejores resultados con un menor número de neuronas de la capa interna. Finalmente, de los clasificadores generados se eligió un clasificador con una eficiencia del 81.25 % usando doce variables de entrada, combinando características de textura y color, y cuatro clases de salida. _______________ IRON AND MANGANESE DEFICIENCIES IN BEAN CROP LEAVES (Phaseolus vulgaris L.) IDENTIFIED BY TEXTURE, COLOR DIGITAL IMAGES ANALYSIS AND ARTIFICIAL NEURAL NETWORKS. ABSTRACT: This research was carried out with the aim to analyze by digitalization bean leaves (Phaseolus vulgaris L.) under different doses of iron (Fe) and manganese (Mn). Eight treatments were set up to evaluate: a control treatment (100% Fe, 100% Mn); two partial deficiencies (50% Fe and 50% Mn); two total deficiencies (0% Fe and 0% Mn), an interaction of the absence of both microelements (0%Fe, 0%Mn) and two doses surplus (200 % Fe y 200 % Mn). The reference for the doses design was the Steiner nutrient solution. An image digital analysis was performed to obtain values from the color spaces RGB and CIELab, plus Chroma (C) and Hue (H) as well as four textural parameters: Angular Second Moment (ASM), Entropy (En), Inertia (IN) and Local Homogeneity (LoH). These color and textural features were extracted from samples of bean leaves (63 days after sowing) treated according with the eight conditions above mentioned. Later, using a different number of variables combination with three data groups: eight, six and four treatments or output variables, various classifiers proposals were generated using neural networks software, the NeuroShell® Classifier Release 2.2 with the aim of classifying the Fe and Mn deficiencies categories. It was found that the use of only textural features results in low accuracy percentages, while the combination of just color features generates better results but it is needed a mayor number of inner layer neurons. But the use of textural and color features generates better correct classifications percentages with a minor number of inner layer neurons. Finally, from the proposed classifiers, one of them was selected, with an 81.25 % of accuracy and the use of twelve input variables involving textural and color features and four output variables.

Master thesis

RGB Textura Redes neuronales Phaseolus vulgaris Hierro Manganeso Texture Neural networks Iron Manganese Edafología Maestría CIENCIAS AGROPECUARIAS Y BIOTECNOLOGÍA

Phenotyping conservation agriculture management effects on ground and aerial remote sensing assessments of maize hybrids performance in Zimbabwe

Adrian Gracia-Romero Omar Vergara Diaz Christian Thierfelder Jill Cairns Shawn Kefauver Jose Luis Araus (2018)

In the coming decades, Sub-Saharan Africa (SSA) faces challenges to sustainably increase food production while keeping pace with continued population growth. Conservation agriculture (CA) has been proposed to enhance soil health and productivity to respond to this situation. Maize is the main staple food in SSA. To increase maize yields, the selection of suitable genotypes and management practices for CA conditions has been explored using remote sensing tools. They may play a fundamental role towards overcoming the traditional limitations of data collection and processing in large scale phenotyping studies. We present the result of a study in which Red-Green-Blue (RGB) and multispectral indexes were evaluated for assessing maize performance under conventional ploughing (CP) and CA practices. Eight hybrids under different planting densities and tillage practices were tested. The measurements were conducted on seedlings at ground level (0.8 m) and from an unmanned aerial vehicle (UAV) platform (30 m), causing a platform proximity effect on the images resolution that did not have any negative impact on the performance of the indexes. Most of the calculated indexes (Green Area (GA) and Normalized Difference Vegetation Index (NDVI)) were significantly affected by tillage conditions increasing their values from CP to CA. Indexes derived from the RGB-images related to canopy greenness performed better at assessing yield differences, potentially due to the greater resolution of the RGB compared with the multispectral data, although this performance was more precise for CP than CA. The correlations of the multispectral indexes with yield were improved by applying a soil-mask derived from a NDVI threshold with the aim of corresponding pixels with vegetation. The results of this study highlight the applicability of remote sensing approaches based on RGB images to the assessment of crop performance and hybrid choice.

Article

Food security Hybrids Maize UAV RGB Aerial Surveying Multispectral MAIZE REMOTE SENSING MULTISPECTRAL IMAGERY CONSERVATION AGRICULTURE CIENCIAS AGROPECUARIAS Y BIOTECNOLOGÍA

Conjunto de componentes para aplicaciones de advertencia en colisión de vehículos

ERIKA PEREZ AGUILAR (2011)

En este trabajo se presenta un método sencillo para la detección de vehículos al frente con el fin de emitir una alarma para la posible realización de una acción.

El método propuesto se basa en la detección del área oscura que se forma debajo del vehículo por la sombra reflejada sobre el pavimento. Para lograr esta detección se seleccionan las áreas de interés, se realzan las sombras o colores oscuros en estas áreas y la información obtenida se convierte a una matriz sencilla donde se hace la búsqueda de los bordes de sombra. Una vez detectado el vehículo de interés en alguna de las zonas de riesgo, se emite una alarma para que el sistema CWS pueda tomar una decisión de acuerdo con la situación.

Los resultados obtenidos demuestran una eficacia del 95%, lo que hace que este método sea competitivo con los métodos que existen actualmente en el mercado.

Master thesis

Security Vehicles Images CIENCIAS FÍSICO MATEMÁTICAS Y CIENCIAS DE LA TIERRA FÍSICA ELECTRÓNICA

Reconstrucción 3D precisa de objetos utilizando un solo sensor Kinect

Precise 3D reconstruction of objects using a single Kinect sensor

MARCELO LUIS RUIZ RODRIGUEZ (2019)

La reconstrucción 3D es un problema en procesamiento de imágenes con múltiples aplicaciones en áreas como medicina, robótica, seguridad, entretenimiento, entre otras. Consiste en generar un modelo 3D de un objeto de interés capturado por un o múltiples sensores, en diferentes tiempos o puntos de vista. Existe un tipo de sensor llamado cámara RGB-D que provee dos tipos de imagen, la información de color y la información de profundidad. La reconstrucción 3D se convierte en un problema complejo en el momento en que el objeto se deforma en el proceso de captura. El presente trabajo de tesis propone un sistema para la reconstrucción 3D utilizando la información geométrica del objeto, adquirida a partir de la información de profundidad, y la información visual del objeto, adquirida a partir de la imagen de color. El sistema se basa en el algoritmo del punto iterativo más cercano (ICP) el cual se encarga de minimizar la distancia euclidiana entre dos pares de nubes de puntos. El problema del algoritmo ICP es que necesita una estimación inicial cercana. Este problema se solucionó utilizando detectores y descriptores, tanto de nubes de puntos como de imágenes. Finalmente el sistema realiza correcciones no rígidas para depurar el proceso de registro.

3D reconstruction is a problem in image processing with multiple applications in areas such as medicine, robotics, security, entertainment, among others. It consists of generating a 3D model of an object of interest captured by one or more sensors, at different times and points of view. There is a type of sensor called RGB-D camera that provides two kinds of image; that is, color information and depth information. 3D reconstruction becomes a challenging problem when the object is deformed during the capture process. This thesis proposes a precise 3D reconstruction system based on using the geometric information of the object, acquired from the depth information, and the visual information of the object, acquired from the color image. The system uses on the algorithm of the iterative closest point (ICP), which minimizes the Euclidean distance between two pairs of point clouds. The problem with the ICP algorithm is that it needs a close initial estimation. This problem was solved by using a combination of detectors and descriptors for both point clouds and images. Finally the system performs non-rigid corrections to improve the registration process.

Master thesis

Reconstrucción 3D, registro, ICP, objetos deformables, cámara RGB-D, Kinect 3D reconstruction, register, ICP, non-rigid objects, RGB-D camera, Kinect INGENIERÍA Y TECNOLOGÍA CIENCIAS TECNOLÓGICAS TECNOLOGÍA DE LOS ORDENADORES INFORMÁTICA INFORMÁTICA

Diseño de algoritmos adaptativos para sistemas SLAM con cámara RGB-D

Design of adaptive algorithms for SLAM systems with RGB-D camera

ANTONIO DE JESUS ORTIZ GONZALEZ (2019)

La localización y el mapeo simultáneo (SLAM) visual es un problema de investigación muy activo en las áreas de la robótica móvil autónoma y visión por computadora, donde un robot necesita localizarse en entornos desconocidos procesando la información de cámaras a bordo sin sistemas de referencia externos como el Sistema de Posicionamiento Global (GPS). En este trabajo, se presenta un SLAM visual basado en características visuales, que es capaz de producir mapas tridimensionales de alta calidad en tiempo real con una cámara RGB-D de bajo costo como el Microsoft Kinect. El mapa generado es adecuado para la planificación futura o tareas comunes de navegación de robots. Primero, se presenta una evaluación integral del rendimiento de la combinación de diferentes detectores y descriptores de características visuales más usados en SLAM. El propósito principal detrás de estas evaluaciones es determinar la mejor combinación de algoritmo de descriptor-detector para usar en la navegación del robot. En segundo lugar, utilizamos el algoritmo RANSAC en combinación con el algoritmo del punto más cercano iterativo (ICP) para obtener el movimiento relativo entre cuadros consecutivos y luego refinar la estimación de pose siguiendo la regla de composición. Sin embargo, la distribución espacial y la resolución de los datos de profundidad afectan el rendimiento de la reconstrucción de escenas 3D basada en RANSAC e ICP. Debido a esto, proponemos una arquitectura adaptativa que calcule la estimación de pose a partir de las mediciones más confiables en un entorno dado. Evaluamos nuestro enfoque ampliamente en los conjuntos de datos de referencia disponibles comúnmente usados en la literatura. Los resultados experimentales demuestran que nuestro sistema puede lidiar robustamente con escenarios desafiantes mientras es lo suficientemente rápido para aplicaciones en línea.

Visual Simultaneous localization and mapping (SLAM) is a very active research problem in the fields of autonomous mobile robotics and computer vision, where a robot needs to localize itself in unknown environments by processing onboard camera information without external referencing systems such as Global Positioning System (GPS). In this work, we present a feature-based visual SLAM, which is able to produce highquality tridimensional maps in real time with a low-cost RGB-D camera such as the Microsoft Kinect. It is suitable for future planning or common robot navigation tasks. First, a comprehensive performance evaluation of the combination of different stateof-the-art feature detectors and descriptors is presented. The main purpose behind these evaluations is to determine the best detector-descriptor combination to use for robot navigation. Second, we use the RANSAC algorithm in combination with the Iterative Closest Point (ICP) algorithm to get the relative motion between consecutive frames and then to refine the pose estimate following the composition rule. However, the spatial distribution and resolution of depth data affect the performance of 3D scene reconstruction based on the RANSAC and ICP. Due to this, we propose an adaptive architecture, which computes the pose estimate from the most reliable measurements in a given environment. We evaluate our approach extensively on common available benchmark datasets. The experimental results demonstrate that our system can robustly deal with challenging scenarios while being fast enough for online applications.

Master thesis

Localización, mapeo, SLAM, cámara RGB-D, característica visual, algoritmo adaptativo, RANSAC, ICP Localization, mapping, SLAM, RGB-D camera, visual feature, adaptive algorithm, RANSAC, ICP INGENIERÍA Y TECNOLOGÍA CIENCIAS TECNOLÓGICAS TECNOLOGÍA DE LOS ORDENADORES INTELIGENCIA ARTIFICIAL INTELIGENCIA ARTIFICIAL

A spatial framework for ex-ante impact assessment of agricultural technologies

Juan Ignacio Rattalino Edreira Peter Craufurd Jordan Chamberlin Lieven Claessens Julius Adewopo Martin van Ittersum Kenneth Cassman (2019)

Traditional agricultural research and extension relies on replicated field experiments, on-farm trials, and demonstration plots to evaluate and adapt agronomic technologies that aim to increase productivity, reduce risk, and protect the environment for a given biophysical and socio-economic context. To date, these efforts lack a generic and robust spatial framework for ex-ante assessment that: (i) provides strategic insight to guide decisions about the number and location of testing sites, (ii) define the target domain for scaling-out a given technology or technology package, and (iii) estimate potential impact from widespread adoption of the technology(ies) being evaluated. In this study, we developed a data-rich spatial framework to guide agricultural research and development (AR&D) prioritization and to perform ex-ante impact assessment. The framework uses “technology extrapolation domains”, which delineate regions with similar climate and soil type combined with other biophysical and socio-economic factors that influence technology adoption. We provide proof of concept for the framework using a maize agronomy project in three sub-Saharan Africa countries (Ethiopia, Nigeria, and Tanzania) as a case study. We used maize area and rural population coverage as indicators to estimate potential project impact in each country. The project conducted 496 nutrient omission trials located at both on-farm and research station sites across these three countries. Reallocation of test sites towards domains with a larger proportion of national maize area could increase coverage of maize area by 79–134% and of rural population by 14–33% in Nigeria and Ethiopia. This study represents a first step in developing a generic, transparent, and scientifically robust framework to estimate ex-ante impact of AR&D programs that aim to increase food production and reduce poverty and hunger.

Article

Agricultural research systems Impact assessment Food production Agricultural R & D Spatial Framework Scaling Out AGRICULTURAL SCIENCES AND BIOTECHNOLOGY IMPACT ASSESSMENT CIENCIAS AGROPECUARIAS Y BIOTECNOLOGÍA