Computer-vision car park monitor

An automated computer-vision tool for monitoring car park occupancy efficiently — combining classical image processing with a YOLO neural network to reach 91% accuracy.

Read the final published paper →

The problem

Prolonged parking overstays result in revenue loss for the parking company.
The current manual method of detecting parking occupancy is inefficient and requires parking officers.

Solution

Implement a high-mounted camera system with computer vision algorithms and deep learning to validate parking spot occupancy.

Key steps

Grey scaling and Gaussian blur
Background subtraction
Thresholding
Dilation
Non-zero pixels
YOLO neural network

Stages of the car park computer vision pipeline: assigning parking spots, background subtraction, thresholding, and final detection output

Results

Method	Accuracy	Explanation
Background subtraction only	50%	Detected every small movement
Background subtraction + thresholding	72%	Filtered out small movements
Background subtraction + Gaussian blur + thresholding + dilation	84%	Reduced noise from the image; still detected humans and other non-vehicle related movements
Addition of YOLO neural network	91%	Eliminated non-vehicle related movement

The final achieved accuracy was 91%, surpassing previous research on this application (IEEE reference).

Future work

Further train YOLO to improve accuracy in classifying detected movements.
Implement multiple cameras to capture images from different angles, preventing small cars from being hidden by larger neighbouring cars.
Develop a web app for parking companies to automate monitoring of their parking spots.