Multi-camera architecture deploys interconnected cameras for comprehensive area coverage. Re-identification is a vital technique within this setup, using feature extraction and matching algorithms to track individuals or objects across varying camera views, crucial for security.
Simulation were created for this project and the results extracted were from this render. Other relavent information were extracted using python script in blender file.
YOLO was used to detect person from frame to frame
If persons identified are within a threshold proximity and the ORB similarity is above a threshold (Step 4) then both persons are considered as one.
The way of feature extract is from the usecase of lookup table approach. Lookup table include - All the position of a person to exist with in a frame
Features of lookup table
- Distance (Person to camera)
- x Axis location of the person (on ground level)
- y Axis location of the person (on ground level)
Distance of the person is found using the formula - ((Height of person * focal lenght) / Bounding box height) After finding distance x axis Location and y axis Location is matched with relavent distance and positioning from the Lookup Table
ORB similarities from opencv library was utilised.
Each person is verified from the saved image folder of people that used to be in the frame for a threshold amount of time.
If the similarities lies above the threshold value then the person is reidentfied from the saved person photo and new image of the person is saved to be refered later.
If the similarities lies below the threshold value then the person is identified as a new person and thier image is saved to be refered later.
Relavent Visualization and Demonstration is provided in this repository in "Final render dashboard" and rest of the .ipynb file.