基于空间序列级联的停车场场景语义分割算法

Semantic Segmentation Algorithm for Parking Scene based on Spatial Sequence Cascade

通过利用深度学习模型对停车场场景中的视觉信息进行处理,以辅助无人驾驶车辆进行自主泊车定位.针对目前用于停车场场景进行语义分割的方法不够成熟,FCN等方法不能满足自主泊车的定位需求,提出一种基于空间序列级联机制的语义分割方法,在停车场场景中,背景及车位线等标识物具有不同的像素区分难度,采用空间序列级联机制,使用浅层网络处理简单像素,使用深层网络处理较难像素,最后将两次结果进行叠加.采集了北京及上海等地共八个不同类型的停车场数据,人工标注40 000张图像构建了停车场场景数据集进行验证,实验表明,在分割精度方面,相比FCN方法,mIoU由69.3%提升到77.9%,速度由15.8 fps提升至35.7 fps.

The visual information in the parking lot is processed by using a deep learning model to assist the unmanned vehicle to locate.Aiming at the current semantic segmentation method for parking lot scenes,the semantic segmentation method based on spatial sequence cascade mechanism is proposed.Eight different types of parking lot data in Beijing and Shanghai are collected and manually labeled.40,000 images were constructed to verify the parking lot scene dataset.Experiment results show that in terms of segmentation accuracy,the mIoU is increased from 69.3%to 77.9%,and the speed is increased from 15.8 fps to 35.7 fps.