-
Murillo Bouzon (PhD Student)
Area: Computer Vision Applied to Medical Area
Title: Automatic 3D Trajectory Planning for Ureteroscopy Based on Meta-Heuristic Optimization and Image Segmentation
Surgeries for the treatment of ureteral stones, also called ureterolithiasis, require the use of a flexible ureteroscope that allows the surgeon to visualize the internal region of the patient's urinary system. For these surgeries, the doctor manually performs a trajectory that starts from the urethra to the region of the ureter, where the stone is located, to then remove it...
However, with the variation of the stone's circumstances, this type of surgery can become complex due to the urinary system having bifurcations and noise, making the trajectory not easy to find, depending on the surgeon's level of experience. The use of robotic surgeries could assist in the treatment, reducing surgery time and risks of patient problems. However, in robotic surgeries, the automatic generation of this trajectory is a fundamental preliminary step for the surgical robot to know the internal environment and reach the destination point without colliding with the patient's internal organs. Automatic 3D trajectory planning can be done using optimization techniques and search algorithms based on the 3D reconstruction of the patient's preoperative computed tomography exams. In addition, immersive visualization of this trajectory can assist less experienced students in better understanding the internal structure of the urinary system and improving their surgical skills by watching the ideal surgical trajectory that should be performed. With this in mind, this work aims to develop a method for automatic planning and immersive visualization of 3D trajectories in ureteroscopic surgeries for the treatment of ureterolithiasis using optimization techniques, 3D space searches, and virtual reality, from preoperative computed tomography exams.
more
-
Fernando Pujaico Rivera (PhD Student)
Area: Computer Vision Applied to Medical Area
Title: Classification of Bedridden Patients' Body Language Using Deep Learning
Human communication involves not only words but also non-verbal elements such as gestures, facial expressions, and body posture. In hospital contexts, the analysis of these interactions is especially relevant, as the early detection of abnormal behaviors in patients can prevent the worsening of serious clinical conditions...
However, automatically detecting the body language of bedridden patients presents significant challenges due to variations in posture, facial expressions, and restricted movements in bed. In addition, the development of robust body language detection systems faces difficulties related to the variety and diversity of information sources, as well as collection conditions in real environments. The computational problem addressed in this research consists of designing an automatic body language classification system using images of bedridden patients, considering image angles from different perspectives, for accurate categorizations. It is proposed to create two databases, BER2024 with simulated images and PER2024 with real images, in addition to a multimodal classification system using deep learning. This system employs convolutional and fully connected neural networks to extract features and classify body language from digital images from different perspectives. The system analyzes facial expression, posture, and full-body images to identify four distinct states: pain, negative body language, positive, and neutral. The results demonstrate the effectiveness of this approach with accuracies reaching 96.47% and 92.51% in the BER2024 and BER2024 + PER2024 databases, respectively. Thus, the high accuracy values obtained, even when the information in the databases used increases in diversity, demonstrate the robustness of the proposed system and indicate promising directions for increasing studies with more variable data.
more
-
Weverson Pereira (Master Student)
Area: Computer Vision Applied to Industry
Title: Computer Vision and Deep Learning for 3D Mesh Alignment in Assembly Assistance in Factory Environments
Equipment assembly is a production process frequently performed in industrial environments, consisting of the joining of individual components or parts to create products that meet the required specifications and satisfy quality standards...
Unlike other industrial processes, which have been automated over the years, much of the assembly process is still performed manually, requiring experienced and trained employees. For this, workers have two main resources: training with experts and a detailed instruction manual outlining the steps to be followed. In both cases, time is required, either for the training of new people or to convert the content of the document into effective operational actions. However, with the advancement of Industry 4.0, the use of immersive realities to assist in the execution of complex tasks is gaining more and more space. Augmented reality, for example, assisted by machine learning for object detection, is capable of converting complex assembly instructions into demonstrative steps, from the overlay of 3D model meshes directly on the workbench. In addition, with the advancement of deep learning, object pose estimation has been widely explored in controlled environments, involving a selected number of objects. However, this task becomes especially challenging in unstructured scenarios, such as in factory environments, which involve tracking multiple objects, where assembly states change and occlusions occur. Thus, the present work proposes a method to assist in the assembly of equipment in unstructured environments, using augmented reality and deep learning. This approach consists of the use of convolutional neural networks for the automatic identification of the position of objects in the scene, as well as the use of recurrent neural networks for the validation of the assembly state at each previously defined stage.
more
-
Andrias Matheus Dias de Pauda, Rafael Gonçalves Monteiro Machado (Undergraduate Thesis)
Area: Computer Vision Applied to Industrial Area
Title: A.R. Detection: A Methodology for Detecting Unsafe Actions on the Factory Floor Based on Deep Learning and Computer Vision
The detection of human activities is one of the most requested problems in the industrial area, especially in places of greater danger such as the factory floor...
With the advancement of the area of computer vision and deep neural networks, the development of methodologies for tracking human actions has been boosted, especially by methodologies based on joint detection. On the other hand, tracking the interactivity of humans with objects, especially in scenes involving a high number of visual noise, the so-called unstructured environments, is an essential task for the prevention of accidents, which are generally common in factory floors, but still not widely addressed in the area of computer vision. Therefore, this project aims to propose and implement a methodology based on deep learning and skeletonization, which tracks both human actions and objects in the scene to measure interactivities and predict the safety of workers, especially on factory floors.
more
