Santiago Sánchez is a student in the Product Design Engineering program at EAFIT University, Medellin, Colombia, where he was born. With his emphasis in mechatronic design, he has developed several projects and gained a lot of experience in modeling, simulation, mechanizing, prototyping and programming. He describes himself as an enthusiastic, passionate and curious person, and his main interest is everything about computer science, artificial intelligence and automation. He also enjoys teaching; that is why he works as a mentor at his university in a subject called Project 7, where he helps students to build an electric manned or unmanned vehicle. Sánchez says that teaching is where he has learned the most. During his free time, he enjoys reading, programming, practicing soccer and extreme sports like kite surfing or paragliding.
Project: Reading Analog Clocks and Gauges
Goal of the project:
Being able to teach a computer to read analog clocks or gauges as a solution for monitoring old machinery and non-electrical equipment like gas tanks, or even for people with visual disabilities.
Summary of work:
First Goal: Reading a gauge
- Train a Predict model (Method→"NearestNeighbors") with 3,000 gauges of eight different types.
Challenges: Ranges and kinds of gauges can vary too much.
Solution: Set one specific kind of gauge and display the range as a percentage to be fitted to any range.
Second Goal: Reading a clock
- Extract the arms of the clock using the "Hough Transform" feature-extraction technique.
- Make a rotating mask over the clock's image, counting the number of pixels in every angle of the mask.
- Train a Predict model (Method→"NearestNeighbors") with 10,000 clocks of six different types.
Challenges: Three arms and the symmetry of clocks were difficult to overcome.
Solution: Specific image processing and a large amount of training data.
Results and future work:
First Goal: 98% accuracy on randomly generated data
Second Goal: 86% accuracy on randomly generated data
For both of the goals, the data used was generated by Wolfram Language functions and some random images from the internet were used, giving less accurate results. It is necessary to train the machine learning algorithm with real labeled images for using it in real applications.