I chose mobile robots because they have many applications and programming them is a difficult task, especially when it comes to detecting objects using cameras and mapping. The robot has many peripheral devices, such as motors and sensors, and can be expanded with an additional arm for lifting objects.
Additionally, this is the first step if you want to design your own ADAS, such as those used in modern cars.
In my latest project, I successfully built a mobile robot integrated with a camera and powered by a Raspberry Pi 4. Utilizing the MobileNetV2 model, I retrained the final layers to recognize specific command signs: "GO," "Turn Right," and "STOP." Leveraging my knowledge of Deep Learning and Transfer Learning, I effectively programmed the robot using Python to interpret these signs and respond accordingly. The robot initiates movement upon seeing the "GO" sign, navigates a right turn when it detects the "Turn Right" sign, and comes to a halt when presented with the "STOP" sign. This demonstration underscores the potential of combining advanced machine learning techniques with practical robotics to create responsive and intelligent systems. Check out the video to see the robot in action!
I became interested in building robots in 2008. These were completely different times, almost no one had heard of Arduino, and Raspberry Pi had not yet been born. Building a mobile robot required more work than today. For the first project, I decided to buy a development board with an Atmel chip. I had got online a books for programming microcontrollers in BASCOM. After a few weeks, I built a six-wheeled vehicle. It had two DC motors and four servos to turn the wheels. There is a proximity sensor installed at the front of the vehicle, which I built from scratch.
In 2009 bought an Arduino Duemilanove which had an ATmega328p on board, a chip with more outputs and inputs than the ATmega8 that was used in six wheeled vehicle. Experimented a bit with diodes and buttons and started building a line follower. Decided that would postpone the six-wheeler project as it required more work and money(let's not forget the crisis). The driving direction is changed by adjusting the speed of the driven wheels. On the front is a sensor module for color recognition. On the back is an arduino and a motor controller built and a color sensor controller.
Until 2011, in my spare time, experimented with sonar and obstacle sensors and a sumo robot, but due to lack of time, stopped doing mobile robot and returned to the topic in 2020. Started building a mobile robot, this time 4 wheeled with Raspberry Pi.
I realized very quickly that in order to continue this project, I had to learn Python. In two month did "Complete Python Developer: Zero to Mastery" and then "Master Computer Vision System and OpenCV3" on Udemy. These courses allowed me to start playing with Raspberry and building a robot that recognizes objects.
As it turned out, the Raspberry Pi 3 has a weak processor and not enough RAM, which makes object recognition very slow. Started delving into the field of object recognition and found an interesting topic like Tensorflow. Decided to use Tensorflow Lite and the built-in model on a Raspberry Pi, but the image was still analyzed at 2 frames per second. That's why I bought a Google Coral accelerator which accelerated object recognition to 8 frames per second.
In 2020, I decided to take the Software Developer course, so I didn't have time to continue with the project until summer 2021.
During the summer break, I used the new skills I acquired in the course and created a web application in Python and Flask
This application allows me to remotely control the robot using a laptop or smartphone.
In 2022, during the Christmas break, I had some time to teach the Raspberry pi to recognize objects.
I trained my system to recognize the remote control, cup, baubles and bin. To do this, I followed instructions from the official tensorflow manual. Prepared 300 photos of objects that I want to recognize. The problem is that for training purposes the size of the photos should be smaller than phone's camera. Photos should not be larger than 800x600 pixels, otherwise more computational time is needed for training. I used the Photoshop to do this, but unfortunately after some time realized it could take hours so had to find another way.
Decided to make my own mini application that changes the resolution of photos. A program that allows me to change the resolution of photos in seconds is available on GitHub. Application allow to reduce size of the photos and videos. You can just download this app, no need to install and add the photos and videos that you want to convert to smaller size and run application.
In the following videos is vision system with Raspberry pi4 (Tensorflow). Using the Coral accelerator allows increase to 14 frames per second.
