Maicol's Dev Lab (TalkinEyes, Mayllo & Web Projects)
- 17 Devlogs
- 31 Total hours
Maicol’s Dev Lab (TalkinEyes, Mayllo & Web Projects)
Maicol’s Dev Lab (TalkinEyes, Mayllo & Web Projects)
Holaaa no supe como crear otros proyectos y enlazar los a haka time jeje entonces aquí pondré los links de los otros dos proyectos jeje
Esta vez nos metimos con matematicass algo extrañas para poder ubicar si el usuario se encuntra con la cara volteada o asi para ajustar las mediciones y ser mas precisas
Devlog: TalkinEyes OS — Días 1 y 2
Objetivo: Evolucionar TalkinEyes V1 (prototipo espagueti) a TalkinEyes OS (arquitectura modular en POO) para el OpenCV AI Competition 2026.
Se ve un poco mas fea que la anterior pero es que el enfoque ahorita esta en hacerlo robusto no tanto bonito
El mayor logro de estos dos días fue reescribir la lógica usando Programación Orientada a Objetos (POO).
En V1, teníamos el mismo código repetido 5 veces para los menús. Ahora, una sola clase Menu maneja todos los menús de forma dinámica.
Se separaron las responsabilidades: la lógica de visión vive en core/, la comunicación en communication/, la interfaz en ui/, y el cerebro en main.py.
El Núcleo de Visión (Modular y Estable)
Se creó un pipeline de visión paso a paso, donde cada archivo hace una sola cosa:
Camera: Abre y gestiona la webcam de forma limpia.
FaceDetector: Envuelve MediaPipe y convierte los frames de BGR a RGB (solucionando errores de memoria contigua).
Creamos una interfaz en Tkinter capaz de cambiar entre formato “cuadrícula” (menú principal) y “lista” (submenús).
Botón de Calibración: Se integró un flujo en vivo que permite al usuario calibrar el sistema sin cerrar la aplicación.
Se dejaron atrás los umbrales universales. Ahora, el sistema mide el EAR del usuario con los ojos abiertos y cerrados.
Calcula umbrales personalizados mediante interpolación lineal.
Persistencia: Los datos se guardan en un archivo perfil_usuario.json y se cargan automáticamente al iniciar la app.
(Debugging)
Manejo de errores de future imports.
Solución al error de MediaPipe: Reference mode is unavailable if data is not c_contiguous.
Protección del sistema contra frames donde no se detecta rostro (evitando cierres inesperados).
Estos dias no hice mucho solo trate de mejorar las calibraciones y hacer unas series de pruebas para que la calibracion tambien ayude aunque la cabeza este inclianda o asi
hemos intentado hacer que maylo abra aplicaciones coom Vscode word Exel etc
Ahora comenzamos con nuevas mejoras para talkin eyes ahora trataremosn de incluir calibracion para disminuir lo mayor posible los falsos positivos
ahora ya que cree la pagina para mi mama comenzare a profundizar mas con lo basico de las paginas web
asi va la pagina de rergalo para mi mamá mañana se la enviare que opinan se ve bien?
It’s been a while, but here’s a new devlog. I have almost the entire structure built and working; there are still some details to iron out, but we’ll adjust those as we go. For now, it speaks, responds, connects to an API, and opens Spotify.
mis primeros pasos en las paginas web
Well, to be honest, I don’t know how to create a new project, so I’ll just continue right here with my project called MYLLO, which is a multimodal assistant type jarvis.
This was my final edition of TalkinEyes; it has already been submitted for grading, so we won’t hear anything until later.
Today, I worked on the main menu interface for TalkingEyes.
I designed the main buttons: EAT, BATHROOM, PAIN, and SLEEP.
I adjusted the sizes, colors, and layout to make the options easier to see.
I continued using Tkinter to build the interface.
I set up the structure so I can add submenus for each option later on.
I didn’t make much progress on the detection side today, but I did make the V1 interface much clearer.
Today, I continued working on the TalkingEyes interface, specifically the blink-based selection system.
First, I fixed a major issue: the block being selected didn’t always match the one visually highlighted. The problem lay in how the index was updated after switching blocks. I adjusted the logic so that the index always represented the currently selected block.
Afterward, I tested the system again and confirmed that the selection process is now working correctly.
I also began implementing the EAT submenu. The idea is that, after selecting a category, various preset phrases appear that can then be selected using blinks.
I started by taking EAR measurements:
Open-eye EAR averaged between 0.20 and 0.30.
Closed-eye EAR was almost always around 0.05 (or similar) down to 0. We also implemented a state machine. Furthermore, we successfully set it up so that an eyelid movement selects a menu item, and both the menu and the camera view are visible.
Devlog: Selection menu + repository setup
Today, we made progress on two fronts. On the interface side, I guided my brother (14, with no prior coding experience) to build the communication menu using Tkinter: four Kahoot-style blocks (EAT, BATHROOM, PAIN, HELP) drawn on a Canvas, featuring an automatic “scanning” highlight system that lights up each block in turn every second—the standard pattern for blink-based accessibility systems.
On the computer vision side, I added frame-counting logic to distinguish an intentional blink from noise or natural blinking, using an EAR threshold and a minimum number of consecutive frames.
I also set up the GitHub repository (though I still need to clean up the .gitignore file to exclude the virtual environment).
Next step: connect both parts so that an intentional blink signals the menu to make a selection.
TalkinEyes Devlog 1 — Starting TalkinEyes
So this is TalkinEyes — a device that lets people who can’t speak or move communicate just by blinking. The idea is the camera watches your eyes, you blink to pick an option from a menu (eat, bathroom, pain, etc.), and it turns that into speech. Kind of inspired by how Stephen Hawking communicated, but way simpler.
I’m building it for a physics/tech event at my university, but honestly I want to keep pushing it further than just that.
What I’ve done so far:
Got face tracking working with MediaPipe — it tracks 468 points on your face in real time through the webcam. Took some fighting with Python/VS versions to get it installed right.
Narrowed it down to just the 6 points per eye I actually need to tell if an eye is open or closed.
Wrote the function that calculates the EAR (basically a number that tells you how open your eye is). Found and fixed a bug where it was only running once instead of updating every frame — that one took me a bit to catch.
Tested it on myself, got EAR values around 0.28 with eyes open, which is right where it should be.
Started keeping a formulas.md file in the repo so it’s clear which parts are formulas/techniques I looked up vs. stuff I built mysel