Abstract: In the process of manually annotating parking slot detection datasets, every effort is made to accurately mark the corner points of parking slot. However, in actual annotation images, some corner points may be blurry or obstructed by obstacles, making precise annotation difficult. This poses challenges for the model when regressing the coordinates of these corner points using the dataset. This article proposes a novel loss function for parking slot corner point offset, which simultaneously learns the corner point offset and localization variance, thereby improving the accuracy of parking space corner point localization with almost no additional computational load. On a self-constructed dataset, the proposed approach achieves higher precision, while significantly reducing both the parking slot corner localization error and the parking slot boundary angle error. Moreover, the standard deviations of these two errors are notably smaller, indicating that the model produces more stable and consistent predictions.

Key words: deep learning; parking slot detection method; Gaussian distribution

摘要： 人工标注的车位检测数据集在标注过程中，需尽可能精准地标注车位角点。但是在实 际的标注图片中，某些车位角点是模糊的或者被障碍物遮挡，难以精准标注，这使得模型在 使用这些标注数据时回归角点坐标变得困难。文章提出了一种新颖的车位角点偏移量损失函 数，同时学习角点偏移量和定位方差，提高了车位角点定位精度，并且几乎不增加额外的计 算负载。在自建数据集上，相对于角点坐标偏移量回归方法，文章提出的方法精准率得到提 高，车位角点误差和车位边线角度误差降低，并且车位角点误差和车位边线角度误差的标准 差更小，模型预测更加稳定。

关键词: 深度学习；车位检测方法；高斯分布