Thanks to my diverse experiences away from home, I have developed resilience through curiosity, creativity and continuous learning. Here follows a selection of professional and personal projects. Feel free to contact me by email if you would like to know more.
Projects
Table of Contents
Camera Focus
LMI Technologies Inc.
Burnaby, BC, CA
2022
Introduction
LMI Technologies designs and builds 3D computer vision solutions for industrial automation. Here is a short overview of the process to obtain camera position and focus adjustment.
Disclaimer
Due to intellectual property, this is a condensed overview of the project and its results. Many details were intentionally omitted.
Case
Creating a platform to automate the characterization and focus adjustment of stereoscopic sensors.
The following process shows how to adjust focus and determine camera position with a single camera.
Note
The process is done twice for the stereoscopic camera.
Process
The sensor to be aligned is a stereoscopic camera with a projected structured light. The goal is to align the left lens with respect to the right against the reference target. The OpenCV code is available on my GitHub.
- Camera calibration and intrinsic values
- Locating camera in space
The Zhang calibration approach:
Reference target :
In an industrial environment the target must remain visible at all times. Here I have simulated camera movement with respect to the target.
Tilt test
Rotation test
Outcome
The target is positioned at a fixed, known distance from the stereoscopic sensor. The position is determined using the sensor’s intrinsic and extrinsic parameters with respect to the target. And feature sharpness serves as a measure of focus.
Automated Quality Inspection
Addev Profom Inc.
Laval, QC, CA
2019
Introduction
Addev Profom inc specializes in the converting and distribution of high-performance materials such as adhesives, technical materials and chemicals.
Case
Reducing quality inspections on die cut gaskets using automation and computer vision.
The product is a foam based gasket with an adhesive backing, delivered to the client on a silicon coated liner
The die-cutting process does not allow for the removal of excess material at the indicated locations (red circle).
Gasket on liner
However, gaskets for automotive clients cannot be delivered with excess material remaining. A significant amount of time was spent manually inspecting finished products and removing any excess when necessary. Here is the approach taken to reallocate manual labor to more valuable tasks.
The Cell
The design is an aluminium extrusion robotic cell with a conveyor belt.
The 4-axis robot moves to the location if material excess is found on the product.
The vision camera is located on the roof of the cell and is considered the origin of the system.
Detection process
The robot is positioned relative to the camera.
The steps involved to determine if excess material is present.
Calibration & Alignment
- Camera calibration (contrast, white balance)
- Camera XY-alignment with respect to the robot
- Robot world calibration with respect to the camera
Video detection
- Circle detect algorithm within the camera FOV
- Find grey level in circle (255 -> no excess/ 0 -> excess)
- Get circle centre if excess is found
- If excess found, stop the conveyor belt
Camera region of interest
Excess removal
Disclaimer
Each application is different and the tool is designed accordingly. Due to intellectual property, the tooling is not shareable.
With the conveyor belt stopped, the robot goes to the reported location and pokes the excess out.
A 3D printed bracket coupled with a poking system was designed and added to the end of the shaft.
Volumetric Video
Vancouver BC, CA
2024
Introduction
I enjoy playing music and recording my performances. To capture different angles, I usually move a camera to multiple locations. However, this requires precise playing for continuity. This project aims to use a single device to film in 3D, eliminating the need for multiple recordings.
Generate a point cloud stream using a custom stereoscopic camera
Creating a video where the user can change camera perspective. This is achieved by scanning scene with the use of a stereoscopic vision system.
Notes
Code available upon request on GitHub with more explanation.
Setup
Use of two webcams and a constant and uniform lighting and backdrop.
Stereo calibration is performed using a similar code seen at Camera Focus
Depth
The bloc matching algorithm is used to calculated the disparity map.
From the disparity map and the camera intrinsic values we deduce the depth map. The depth map was rendered using Google CoLab accelerated GPUs.
Seen from the left camera, depth map and original video composited.
Details
My computer is very slow and doesn’t have a gpu. The depth map was rendered using Google Colab cloud GPUs.
Point Cloud
From the depth map, I obtain a point scatter with OpenCV 3D re-projection method.
Observation
Bloc matching does not perform well with low contrast features.
Many missing points due to obstruction and low quality cameras.
My face scan
Improved point cloud
Observation
The point cloud is cleaner than the first attempt. Fewer gaps in the point cloud are observed.
More scans
Guitar close up
Distant scan
Next steps
- Write algorithm .ply file stream
- Fill gaps in point cloud
OCR Video
Vancouver, BC, CA
2025
Introduction
Reusing the homography code written for Camera Focus to perform live text recognition from a video.
Live video text recognition with ocr result overlay
Setup
Extracting text from book cover with tesseract.
Detail
Code available on my GitHub
Isolating the book the background and apply polar lines
Result
There’s an underlying bug in the code which creates these glitches. Still working on it. Text recognition algorithm needs tuning
Pose Estimation
Vancouver, BC, CA
2024
Introduction
Using pose estimation to determine movement accuracy.
Realtime pose estimation applied to dance moves.
Setup
Dance moves are repeated twice to establish a comparison. Relative movement is measured. OpenCV and Mediapipe
Detail
Code available on my GitHub
Red lines represent the displacement with respect to the attempted pose (frames on the right).
Graph
Accuracy through time
This is still a work in progress.
Result
Dance moves and relative position
Next
- Find subject centre of gravity
- Adjust for multiple angles
Plant Stop motion
Vancouver, BC, CA
2024
Introduction
This is an IoT project with a raspberry pi and Philipps hue light: using computer vision to autonomously take photos of a growing plant.
Taking pictures of a plant to create a stop motion footage.
Setup
Multiple pictures are taken automatically during the day.
Detail
Code available on my GitHub
Pictures are taken according to local sunrise and sunset times. A bluetooth remote controlled light is turned on after sunset. Minor Image processing and histogram matching is applied.
Result
It was difficult to automate the exposure adjustment due to sunlight coming from all directions in the afternoon. The first frame is the colour balance reference for all the frames.