Elevator button recognition is a critical function to realize the autonomous operation of elevators. However, challenging image conditions and various image distortions make it difficult to recognize buttons accurately. To fill this gap, we propose a novel deep learning-based approach, which aims to autonomously correct perspective distortions of elevator button images based on button corner detection results. First, we leverage a novel image segmentation model and the Hough Transform method to obtain button segmentation and button corner detection results. Then, pixel coordinates of standard button corners are used as reference features to estimate camera motions for correcting perspective distortions. Fifteen elevator button images are captured from different angles of view as the dataset. The experimental results demonstrate that our proposed approach is capable of estimating camera motions and removing perspective distortions of elevator button images with high accuracy.
翻译:电梯按钮识别是实现电梯自动操作的关键功能。 然而, 挑战图像条件和各种图像扭曲使得很难准确识别按钮。 为了填补这一空白, 我们提议了一种新的深层次的学习方法, 目的是根据按钮角检测结果, 自主地纠正电梯按钮图像的扭曲。 首先, 我们利用一个新的图像分割模型和哈夫变换方法获取按钮分割和按钮角检测结果。 然后, 标准按钮角的像素坐标被作为参考特征, 用于估算相机移动以纠正视图扭曲。 15个电梯按钮图像被从不同角度作为数据集捕获。 实验结果显示, 我们提出的方法能够非常精确地估计相机动作并消除电梯按钮图像的扭曲。