Parking Lot Occupancy Detection with Improved MobileNetV3

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Parking lot management systems are increasingly researched, with deep learning showing promise for occupancy detection.
• Accurate vehicle occupancy determination is crucial for efficient parking management.

Purpose of the Study:

• To develop an advanced deep learning model for precise vehicle occupancy detection in individual parking spaces.
• To enhance the MobileNetV3 architecture for improved performance in parking lot management systems.

Main Methods:

• An optimized MobileNetV3 model with custom architectural enhancements was developed.
• The model incorporates a convolutional block attention mechanism and blueprint separable convolutions.
• Training was performed on the CNRPark-EXT and PKLOT datasets using individual parking space patches.

Main Results:

• The enhanced MobileNetV3 achieved an Area Under the ROC Curve (AUC) of 0.99 on the PKLot dataset.
• The model attained an average accuracy of 98.01% on combined datasets, outperforming CarNet (97.03%).
• The proposed model demonstrates superior performance compared to state-of-the-art methods like CarNet and mAlexNet.

Conclusions:

• The enhanced MobileNetV3 model offers a robust and efficient solution for real-time parking occupancy detection.
• This advancement has significant implications for improving urban mobility and optimizing parking resource allocation.
• The study contributes to the field of intelligent transportation systems through improved parking management technology.