Vacant Parking Slot Detection in the Around View Image Based on Deep Learning

Area of Science:

• Computer Vision
• Artificial Intelligence
• Automotive Systems

Background:

• Complex visual environments (lighting variations, shadows) challenge accurate vacant parking slot detection for park assist systems (PAS).
• Standalone around view monitors (AVM) require enhanced accuracy in detecting available parking spaces.

Purpose of the Study:

• To propose a novel deep learning-based method, VPS-Net, for robust vacant parking slot detection.
• To improve the accuracy and reliability of parking assist systems.

Main Methods:

• VPS-Net employs a two-step approach: parking slot detection and occupancy classification.
• Parking slot detection utilizes YOLOv3, integrating classification and geometric cue-based localization of marking points.
• Occupancy classification employs a customized deep learning network with optimized convolution kernel size and layers.

Main Results:

• VPS-Net achieved high performance on the ps2.0 dataset with 99.63% precision and 99.31% recall for vacant parking slot detection.
• The method demonstrated satisfying generalizability on the PSV dataset.
• The combined multi-object detection and classification network ensures robust detection of diverse parking slots.

Conclusions:

• VPS-Net offers a significant improvement in vacant parking slot detection accuracy and robustness.
• The proposed method enhances the capabilities of park assist systems in complex visual environments.
• VPS-Net provides a reliable solution for real-world parking assistance applications.