Vision-based Queue Monitoring System

A vision-based queue monitoring system with YOLO-v3 backbone to count number of people in queues at supermarket, retail stores, train station & airport, hospital, etc.

Comment

Inspiration

When I go to NTUC Fairprice and queue for checkout, the queue is so long and people are so crowded. I want to find a short queue and also near me, but finding one is difficult. It makes queuing is much more time-consuming. Other statistics about queuing:

  • Each person spends about 7 minutes in queuing every week on average. [1]
  • Singapore population is 5,900,000 in 2022 [2]
  • That means, Singaporean waste about 35,000,000 hours a year for queuing

What it does

After user input the regions of queues by drawing on the screen, the system connects to CCTV camera and detects person object in pre-defined regions and counts number of people in each regions, then displays the result in a large monitor.

How we built it

Use a computer vision model (YOLO-v3) to detect human in pre-defined regions --> Develop an algorithm to count number of people in each region that users have chosen --> Run inference of the model on webcam (simulating CCTV camera) --> Apply the people counting algorithm --> Display the result to pygame.

Impact of the system

Average salary per hour of a Singaporean is S$23.36/hour [3], if each Singaporean save 1 minute for queuing each day thanks to our system, Singapore will save S$820,000,000 each year from queuing. Other benefits:

  • Fast process and save time
  • Improve customer experience
  • Increase productivity for both customers and staff
  • Optimise operation cost and efficiency
  • Better communication and space management
  • Allow customer data analysis

Where can our system apply for

  • Supermarket
  • Retail stores
  • Hospital
  • Bank
  • Train station & airport

What's next for Vision-based Queue Monitoring System

In further stages, we will try to estimate the waiting time for each queue. We will also improve the model by adopting 3D object detection (by stereo camera) to detect people and define kiosk regions more accurately. Moreover, setting up more cameras with better angle (top-down angle which we could see human's head and little occlusion) will also improve the accuracy. Furthermore, another model to estimate the number of items each person buy will also be developed to give customers information not only the number of people but also the amount of goods people buy in the queue.

Built With

Share this project:

Updates