Real-Time Parking Space Detection Based on Deep Learning and Panoramic Images.

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Parking space detection is crucial for autonomous parking systems.
• Existing datasets lack diversity for robust real-world performance.
• Accurate real-time detection remains a significant challenge.

Purpose of the Study:

• To develop an improved algorithm for accurate and real-time parking space detection.
• To create a diverse panoramic parking space dataset (PSEX) using Generative Adversarial Network (GAN) image style transfer.
• To enhance the performance of PP-Yoloe for detecting parking space state and angle.

Main Methods:

• Constructed the PSEX dataset by integrating GAN-based image style transfer.
• Modified the PP-Yoloe algorithm by replacing ResSpp with ResSimSppf, introducing SimSppf at the neck, changing activation functions, and adding an auxiliary detector head.
• Evaluated the SimSppf_mepre-Yoloe model on parking space detection tasks.

Main Results:

• The SimSppf_mepre-Yoloe model achieved average improvements of 4.5% in mAP50 and 2.95% in mAP50:95 over the baseline PP-Yoloe.
• The model maintained comparable inference latency, reaching 33.7 FPS on Jetson AGX Xavier with TensorRT optimization.
• GAN-based data augmentation significantly enriched the diversity of parking space data.

Conclusions:

• The proposed model offers a superior balance between detection accuracy and real-time performance for parking space detection.
• The enhanced dataset and algorithm are suitable for deployment in intelligent vehicles and robotic perception systems.
• This research advances the capabilities of automatic parking systems towards full autonomy.