Hackathon- Lift Bot#
Created: October 24, 2021 8:29 PM
Tags: Raspberry Pi
Pain Point Description#
After the company moved, the office space and surrounding environment have been greatly improved. The only thing that has regressed is the elevator in the building. The elevator scheduling algorithm is terrible, and it takes 5-10 minutes to wait for the elevator each time. Each person can spend nearly 20-30 minutes waiting for the elevator every day. The company has about 500 employees, 30 * 500/60 = 250 hours, 5500 hours per month, which is equivalent to wasting 550 days of work time every month. At the same time, when visitors arrive, it is also easy to leave a bad impression. I often think, it would be great if we could save this time.
How to Solve#
Find a person to stand at the elevator entrance. When someone wants to take the elevator, they press the button and notify the person who wants to take the elevator when the elevator arrives.
By using this method, we can still save 5280 hours per month. However, isn't this a bit stupid?
So, we use a computer to do these things for us.
Send a request to the Raspberry Pi through a mobile phone, trigger the servo to press the elevator button, start the vision kit, use deep learning for computer vision to recognize the floor information on the elevator LED display, and send a notification to take the elevator when the elevator is approaching.
Hardware Equipment#
Host#
Raspberry Pi Zero W * 2
Visual Recognition#
Google Vision Kit * 1
Tensorflow deep learning board, camera, button, LED, buzzer
Mechanical Drive#
WS-SG900 servo * 1
Servo drive board * 1
Network Section#
Each Raspberry Pi has started a flask web service to receive and send commands.
Raspberry Pi with Servo#
There is an interface for receiving button commands, which is connected to a servo drive board to receive servo execution commands and drive the servo.
Raspberry Pi with Computer Vision#
It provides an interface for receiving computer vision recognition commands, calls Google's AIY Vision board to recognize the elevator's dashboard, and when it recognizes that the elevator is approaching the position of the 12th floor up or down by 3 positions, it calls the Feishu API to send a message to the corresponding person.
Model Section#
The mobile version of Google Tensorflow's mobilenet recognition model is used for supervised learning classification. It recognizes different floors and going up or down the stairs.
Data Collection#
Collected over 200 images of going up and down different floors. Removed hidden files on Mac.
Labeling#
Divided into 30 different labels, up to 1, down to 1. Each floor is divided into these two categories of images.
Defects#
Due to time constraints, the model parameters have not been well tuned, and the recognition rate is less than 80%.
Results#
Won the first place in the company's 2021 Hackathon and received a prize of 10,000 yuan.
Todo
Post a video of the production process and the effect.