Andrea Pilco

This study presents the development and validation of a real-time optical character recognition (OCR) system for identity verification in residential access control scenarios, specifically evaluating the effects of environmental variability and user interaction on system performance. The proposed methodology integrates a U-Net architecture with a ResNet-34 encoder for semantic document segmentation and utilizes Tesseract OCR for digit recognition, deployed on an embedded platform. Experimental evaluations were performed under diverse lighting conditions, varying camera-to-document distances, and differing document physical states, followed by field validation within a gated community. The system demonstrated optimal segmentation performance with a mean Intersection over Union (IoU) of 0.82 and a digit recognition accuracy of 87.5% under high-illumination conditions at distances ranging from 7 to 30 cm. Furthermore, system usability and user workload were assessed through standardized psychometric instruments, including the System Usability Scale (SUS), the User Experience Questionnaire (UEQ), task-efficiency metrics, and the NASA Task Load Index (NASA-TLX). Results indicated high perceived usability (SUS = 86.5) coupled with moderate cognitive demand. The findings demonstrate that while environmental factors significantly influence algorithmic performance, user interaction critically impacts overall system reliability. The novelty of this research lies in integrating technical performance metrics with human-centered analysis in a real-world deployment.

Welding plays a critical role in industrial manufacturing, defects such as cracks, porosity, and incomplete fusion compromise structural integrity. Traditional inspection techniques, including manual assessment and non-destructive testing (NDT), are labor-intensive and prone to human error. This study presents a comparative study on the visual classification of weld bead defects in Shielded Metal Arc Welding (SMAW) using Convolutional Neural Networks (CNNs). The study used two image datasets: one with image preprocessing techniques applied (sharpening, adaptive contrast adjustment, and Gaussian smoothing), and another without any preprocessing. Both datasets were used to train and evaluate a CNN model under controlled laboratory conditions. The results demonstrate that preprocessing improves model performance, increasing accuracy from 91.66% (without filters) to 98.00% (with filters). A detailed comparison with related works highlights the contribution of this approach, which not only achieves competitive accuracy but also introduces a replicable methodology for dataset preparation and analysis.

Drunk driving remains a leading cause of traffic crashes worldwide. To better understand impaired driving behaviors, this study utilized a driving simulator to compare two scenarios: one where participants consumed alcohol and another where they drove without alcohol's effects. A virtual environment replicating the city of Palmira, Colombia, was created using geospatial data provided by GIS and procedural modeling techniques to enhance realism, including the customization of characteristic buildings. Autonomous non-player characters (NPCs) were integrated using NavMesh agents, and external controllers such as a steering wheel and pedals were employed to replicate real driving conditions. The simulation evaluated driving behaviors, focusing on the blurred vision and delayed steering responses associated with alcohol consumption. Thirty students participated in the study, driving a predefined 3kilometer route. Participants were divided into two groups: those who consumed alcohol and those who did not. A difference of 1 minute and 25 seconds in driving time was observed between the two groups, with alcohol consumption contributing to slower response times and impaired visibility. A post-simulation questionnaire assessed participants' perceptions of the simulation. The environment's realism received an average score of 4.8/5, while the simulation's immersion was rated at 3.7/5. This experiment highlights the potential of virtual simulations in studying impaired driving behaviors.

In this paper, a sign language recognition system for the Ecuadorian Sign Language vowels (A, E, I, O, U) using Random Forest (RF) and YOLOv8 models is proposed. For this purpose, a new dataset with a total of 500 RGB images in natural light for single-hand gestures was created. RF model used the normalized hand landmark coordinates obtained by using Mediapipe while for real-time gesture detection, YOLOv8 took images with higher resolutions. Hypothesis testing results also showed that the RF model had better accuracy, precision, recall, and computational complexity with the accuracy, precision and Recall scores all 100 % and were preferred for real-time applications. YOLOv8 performance was high with a precision of 100% revealing the model as suitable for tasks related to images. Final real-time inference tests validated our claims of scalability and efficiency of RF as it was able to classify gestures within an average of 0.0055 seconds of inference time. This paper underscores the importance of machine learning models in enhancing inclusion as well as closing communication barriers for the hearing-impaired population.