BEV-PolyNet: BEV-Based Polygonal End to End Parking Slot Detection Framework

Abstract

Accurate parking slot detection is crucial for autonomous parking and intelligent driving, directly impacting safety and efficiency. However, most existing methods rely on AVM images, making them susceptible to image distortion and vehicle occlusion. Additionally, many approaches independently regress corner points without considering the overall parking slot structure, limiting detection accuracy and necessitating post-processing. To address these challenges, we propose BEV-PolyNet, a novel end-to-end detection framework. Firstly, we generate BEV features using surround view images to address the distortion and occlusion issues caused by AVM images. Secondly, we introduce a polygonal modeling approach that preserves the integrity of parking slots while accommodating complex shapes. Finally, we improve detection accuracy and convergence speed by initializing query vectors with features from surround-view monitoring images, combined with location prior encoding. Extensive experiments on the public PS2.0 dataset and a private LPD dataset validate the effectiveness of BEV-PolyNet, demonstrating its superior performance in parking slot detection.