一种基于环视图像的空停车位检测方法

A vacant parking slot detection method based on around view images

停车位检测是自动泊车系统的重要组成部分.针对车位线及角点不清晰或出现遮挡情况下的车位难以检测问题,提出了一种基于环视图像的空停车位检测方法.该方法将整个空停车位视为一个整体,在空停车位车位线及角点信息不完整的情况下利用MobileNetV3进行特征提取并构建特征金字塔,进而实现分类、边框回归和语义分割,再根据Douglas-Peucker算法对语义分割结果进行拟合得到空停车位.实验表明,所提方法在ps2.0数据集测试集中挑选出的车位线及角点不清晰或者出现遮挡的197张测试图片中检测性能优秀,精确率达到98.39%,召回率达到97.21%,实现单帧11 ms的检测速度,能对不同环境下的不同类型车位进行检测.此外该方法在自注释数据集上具有优秀的性能,有较好的泛化能力.

Parking slot detection is an essential part of the automatic parking system. Aiming at the problem that parking slot is difficult to detect when the parking slot line and corner points are unclear or occluded, a vacant parking slot detection method based on around view images is proposed. This method treats the entire vacant parking slot as a whole, uses the MobileNetV3 to extract features and build a feature pyramid network when the parking slot line and corner points information of the vacant parking slot is incomplete, and then achieves classification, border regression and semantic segmentation. The Douglas-Peucker algorithm is adopted to fit the semantic segmentation results to obtain vacant parking slot. Experiments show that this method has good detection performance in the 197 test pictures with unclear or occluded parking slot line and corner points selected in the test set of ps2.0 dataset, with an accuracy rate of 98.39% and a recall rate of 97.21%, and a single frame detection speed of 11 ms is realized, it can detect different types of parking slots in different environments. Moreover, it has good performance and satisfying generalization ability on the self-annotated dataset.