Technical Independent Study
Application of computer vision model based on yolov8 in acquiring and processing human flow data in MTR
by Wuzi
01
YOLOv8 model
We used the yolov8 model, which is a powerful CNN-based visual recognition model open-sourced by ultralytics. It can directly perform basic object detection and classification recognition. At the same time, we can also develop many derivative functions as needed.
02
Object Detection and Recognition
The basic yolov8 function can identify all common objects within the camera range, frame the identified objects with a box, name them with tags and indicate the confidence level.
03
Object Recognition of a Specific Class &
Object Counting in Regions &
Heat Mape
By giving yolov8 specific instructions, we can let it recognize specific types of objects in specific areas, which is very helpful for us to count MTR traffic data. We can also generate heat maps based on the time the object stays in the picture to visualize flow data.
04
Train Model to Detect Head Only &
Object Blurring
We tried to train our own model. By using the dataset we obtained, we trained the yolov8n model and obtained a model for head recognition. This progress can help us reduce the recognition errors caused by the mutual occlusion of human bodies on the MTR platform. At the same time, we also blurred the detected heads by processing the images frame by frame to protect the privacy of customers.
05
Overall Outcome
Since we could not get a response from MTR, we finally tried running our CV recognition subsystem in McDonald’s and the result was very successful, with a recognition accuracy of over 90%.
