Object ground lines regression and mapping from fisheye images to around view image for the AVP

Abstract

The automated valet parking (AVP) based on the around view monitor (AVM) has attracted increasing attention from the research community due to its high practicality and low cost. However, the distortion correction and inverse perspective transformation of the around view images result in the truncated and deformed objects, which leads to poor detection performance. To address this problem, we propose a method of object ground lines regression and mapping from fisheye images to around view images. There are two key steps for the proposed method, regressing ground lines in the fisheye images and mapping ground lines to the around view image. To detect the objects and regress ground lines at the same time, we explored three new detectors based on YOLOv5 (i.e. YOLOv5-VGG, YOLOv5-Point, and YOLOv5-Ratio). Among them, YOLOv5-VGG adds a VGG-based regression network behind YOLOv5, YOLOv5-Point directly adds a regression head to share the convolutional features, and YOLOv5-Ratio combines the 2D box and ground line to predict its ratios. After that, the ground lines from four different views are mapped and fused to the around view image. Additionally, to facilitate the study of the object detection in the around view image, a large-scale labeled dataset is established, which comprises 9828 fisheye images collected from typical indoor and outdoor parking lots. For each image, the 2D bounding boxes and ground lines of objects are carefully labeled. Experiments show that the YOLOv5-Ratio obtains the best performance and can effectively detect the truncated and deformed object in the around view image with the precision rate of 92.28% and recall rate of 85.09% on our collected dataset.