Miércoles, 15 de noviembre
Session: Human Monitoring
9.00 – 9.45 horas
Title: Deep learning for gait recognition and embedded systems
Francisco Manuel Castro
Universidad de Málaga
Abstract: Nowadays, the most successful approaches for computer vision problems are developed using Deep Learning models like CNNs or Transformers. These models require massive labeled datasets and computational resources to be trained and deployed. However, the amount of training data is very scarce in real-life problems like people identification. Moreover, with the advent of IoT and edge computing, a new trend exists to deploy deep learning models on embedded systems to increase privacy and reduce the system’s latency. In this talk, we will focus on these two topics: (i) designing deep learning models using the gait information as a biometric pattern for people identification; and, (ii) efficient design and deployment of deep learning models on embedded devices.
9.45 – 10.15 horas
Title: Towards Human Action Recognition for Spanish Sign Language
Universidad de La Rioja e IR Soluciones
Abstract: The recognition of human gestures and actions in images, and videos is an active area of research in computer vision. This field has made great advances in the last decade thanks to the use of deep learning techniques. Furthermore, the recent spread of low-cost video camera systems, including depth cameras, has reinforced the development of observation systems in a variety of application areas, such as video surveillance, home security, healthcare, etc. However, most of these developments are in controlled environments. The recognition of movements and gestures in real-time through a camera that acquires its images in an uncontrolled environment (such as a shopping mall, a university lobby, or a museum hall), allows one to interact with passers-by in these scenarios. Dealing with this topic implies challenges in various areas that include at least technological, social, and legal challenges that need a careful approach. Within this framework, we set as objectives of this project the design of a web interface adapted to interaction without physical contact (i.e., through video images captured with a camera in real-time) and the construction of artificial intelligence models (based on deep learning) that guarantee, in an uncontrolled environment, an interaction with this web interface. As a first step in this work, we propose a literature review in this area, the creation of two datasets based on the vowels of Spanish Sign Language (SSL). Afterward, we trained five gesture classification models. We obtained an accuracy of 99.2% and integrated this model with a hand detection model to perform gesture recognition in real-time for the gestures of vowels of the SSL.
10.15 – 11.00 horas
Title: Human Action Segmentation in manufacturing contexts: comparisons among deep learning models trained with multimodal features
Abstract: Industrial robots have become prevalent in manufacturing due to their advantages of accuracy, speed, and reduced operator fatigue. Nevertheless, human operators hold a pivotal position in primary production lines, following the principles of Industry 5.0. To identify the physical and cognitive behavior of operators working alongside collaborative robots, it is fundamental to focus on the temporal segmentation of human actions. This talk delves into the utilization of manufacturing-centric datasets as a key role in developing and training deep learning models for human action segmentation. The talk introduces a novel multimodal dataset and conducts an in-depth examination of various deep learning architectures, such as temporal convolutional networks, graph convolutional networks, and transformers. Such models are trained by splitting the dataset and differentiating from both users and work locations, aiming to study the generalization of the data. The features used for training the models are extracted from RGB, Depth, and Skeletal information within the dataset, also considering combining sets of features to evaluate trainings with multimodal data.
11.00 – 11.30 horas
Session: Applications to medicine
11.30 – 12.15 horas
Title: Towards the Digital Uropathologist: Deep Learning for Histopathological Diagnosis of Prostate Cancer
Instituto de Ingeniería de Aragón
Abstract: Prostate cancer stands as the most prevalent malignancy among men and ranks third in terms of mortality, exhibiting an escalating incidence trend. When detected at an early stage, the prognosis is favorable. However, an overestimation of the tumor’s aggressiveness often leads to aggressive treatment options such as radical prostatectomy, which may entail substantial side effects, including urinary incontinence and erectile dysfunction. To accurately assess the disease’s stage and aggressiveness, biopsies are conducted in patients presenting abnormal biomolecular indicators. This assessment is now standardized through the Gleason score. Nevertheless, it is semiquantitative and bears a significant degree of subjectivity tied to the experience of the uropathologist. In this context, we have employed a database of histologies meticulously annotated by a team of pathologists who analyzed specimens from radical prostatectomies. We have trained a classifier based on the Cross-Covariance Image Transformers architecture to identify distinct histological structures. The network’s inference is complemented by a filtering process to characterize these pathological structures at the histological preparation level.
The trained classifier exhibits a 97% accuracy when evaluated in a test data-set. Preliminary findings in newly acquired images, wherein diverse patterns have been reconstructed, are promising, particularly in well-represented categories. However, it’s essential to emphasize that these results are preliminary and warrant validation. The proposed methodology, when applied to biopsy images, offers a more flexible and less subjective means of identifying various patterns compared to the Gleason score. This approach not only enables a richer quantification but also paves the way for a more precise diagnostic process.
12.15 – 13.00 horas
Title: Advancing the computational assessment of retinal vascular tortuosity as a clinical biomarker
Universidade Da Coruña
Abstract: Retinal vascular tortuosity has a valuable potential as biomarker for many relevant vascular and systemic diseases. However, the lack of a precise and standard guide for tortuosity evaluation and the high inter and intra expert variability hinder an objective and reliable assessment, limiting its use for diagnostic and treatment purposes. This work is intended to establish the basis for advancing in the standardization of the retinal vascular tortuosity as a clinical biomarker with diagnostic potential, allowing, thereby, objective computational measurements that better represent the analysis performed by the specialists in the clinical practice. For this purpose, first it is conducted a multi-expert analysis to assess the maturity and consistency of clinical criteria and to establish a categorization that provides a reliable representation of specialists’ perception. Then, it is presented a computational tortuosity metric that integrates the mathematical representation of vessel tortuosity with anatomical properties of the fundus image identified as relevant by the specialists. Finally, the diagnostic potential of the proposed metric is analyzed for several pathologies. The evaluation of its prognostic performance shows that the integration of the anatomical factors provides an accurate, objective tortuosity assessment aligned with the insights of specialists.
13.00 – 15.00 horas
Session: Precision agriculture
15.00 – 15.45 horas
Title: Robotic technologies for in-field crop monitoring
Abstract: Agricultural robots promise to provide effective solutions to improve task efficiency over large fields. However, the use of farmer robots is still under investigation and several challenges, including accurate vehicle localization and control, need to be addressed to increase autonomy and safety in scarcely structured environments, such as vineyards. One of the main challenges in this respect is to make the robot able to follow a trajectory parallel to the vineyard row, avoiding collisions with the crop. To address this challenge, the developed algorithm exploits sensor data from a cost-effective multi-sensor system, which includes an inertial measurement unit (IMU), a global navigation satellite system (GNSS) receiver, and an RGB-D camera. Data from these sensors are integrated through an information filter and then employed in a control algorithm designed to enable the robot to safely follow a vineyard row. While navigating throughout the crop, the robot is able to gather and process onboard visual and depth data provided by an RGB-D sensing device to extract information on the plant health status, such as vegetation indexes and morphological measurements. All measurements are geo-referenced by an algorithm that synchronizes the positioning information obtained from the robot’s localization system with the camera data. The output of the algorithm is stored in a database from which users can request information regarding a specific survey. The autonomous navigation algorithm is developed under ROS framework and validated in a simulated environment built in Gazebo. The service for farmers, on the other hand, is developed in Flask.
15.45 – 16.30 horas
Title: Towards AI-based Production Estimation and Biomass Quantification of Individal Plants for Precision Agriculture
Abstract: The technological revolution due to the development of Artificial Intelligence (AI) plays a pivotal role in enhancing productivity and sustainability in precision agriculture. This presentation delves into the application of AI-driven methodologies for production estimation in vineyards, tomato plantations, and pomegranate orchards. Several approaches to data management are presented to help fruit detection for yield prediction and fruit quality assessment throughout the estimation of the fruit size. Alongside, this talk explores the application of AI in pruning tasks, elucidating how machine learning algorithms can measure the biomass removed during automated pruning processes. Such advancements not only optimize crop yields but also aid in efficient resource management and waste reduction. Attendees will gain practical insights into the implications of AI, setting the stage for a more sustainable agricultural future.
16.30 – 17.15 horas
Title: Automation and Control in Protected Cultivation: Past, Present and Future Trends
José Boaventura Ribeiro Da Cunha
Universidad de Trás-os-Montes e Alto Douro, Portugal
Abstract: Climate changes and an increasingly rising population highlight the need for sustainable food production solutions, mainly in terms of energy, land and water use. A relevant question that we face today is how to increase production year-round, with the same available land areas and with balanced costs? Growers and researches have devoted their attention to the so-called Controlled Environment Agriculture or Protected Cultivation. Nowadays, Protected Cultivation is a current agriculture practice that uses several hardware and software toolsto change climatic factors such as temperature, light, Humidity, CO2 concentration, nutrient supply, among other factors to achieve optimal plant performance and higher and better quality yields. This talk aims to present the state-of-the-art of the technological developments and applications in protected cultivation. Aspects such as modeling, instrumentation, energy optimization and applied robotics are discussed, aiming at not only to identify latest research topics, but also to foster continuous improvement in key cutting-edge problems. The limitations, advantages and challenges faced by Protected Cultivation technologies are addressed with the aim of identifying the prospective future developments.
Jueves, 16 de noviembre
10.00 -11.30 horas
Title: Deep Detection and Segmentation Models for Plant Physiology and Precision Agriculture
Ángela Casado García
Universidad de La Rioja
Abstract: In this thesis, we have focused on developing methods to improve the performance of deep object detection models. To achieve such a goal, we have used ensemble methods to devise an algorithm that enhances the accuracy and robustness of object detection models. Moreover, the proposed algorithm is the basis for defining semi-supervised learning techniques that reduce the number of annotated images that are required to train object detection models. In addition, we have simplified the creation and use of detection models by building an easy-to-use graphical interface. The developed methods and tools are not only applicable to object detection problems, but we have generalised them to a different computer vision task that is semantic segmentation. Finally, our work is not only theoretical, but it has also been applied to tackle actual problems in plant physiology and precision agriculture.
This event has been partially funded by the «Vicerrectorado de Investigación» of «Universidad de La Rioja» and by Grant PID2020-115225RB-I00 funded by MCIN/AEI/ 10.13039/501100011033. The «Departamento de Matemáticas y Computación» of «Universidad de La Rioja» also has contributed to the event by facilitating the required equipment.
Jónathan Heras Vicente
Vico Pascual Martínez-Losa
Departamento de Matemáticas y Computación
Universidad de La Rioja