摘要
在车载全景系统中,如何准确地检测停车位的位置和车位的方向仍然是一个待解决的问题。针对这一问题,设计了一个双路并行多尺度标记点检测网络,双路网络分别用于检测标记点的位置和角度。对全景图像提取多尺度的特征,并行维护一高一低两个分辨率的分支网络,两个分支互相融合,高分辨率特征以高斯热图的形式表述标记点的位置。提出了一种新的停车位方向计算方法,使用两个标记点的方向以及两个标记点的相对位置计算停车位的方向。为验证所提方法的可行性,使用公共数据集PS2.0的训练集训练所设计的网络,在公共数据集PS2.0和自行采集的数据集PSS上分别测试的停车位检测精确度为99.4%、95.27%,召回率为99.88%、80.89%,在PS2.0上标记点位置平均误差为0.84 pixel,车位方向的误差为0.71°。实验结果表明,与现有方法相比,所提出的车位检测网络降低了标记点定位和车位方向的误差,且在PSS数据集上有较强的泛化能力。
In the vehicle panoramic system,how to accurately detect the position and direction of parking slot is still a problem to be solved.To solve this problem,we designed a two-way parallel multi-scale mark point detection network.The two-way network is used to detect the position and angle of mark point respectively.Multi-scale features are extracted from panoramic images,and a branch network of one high and one low resolution is maintained in parallel.The two branches are fused with each other.The high-resolution features express the location of mark point in the form of Gaussian heatmap.A new method for calculating the direction of parking slots is proposed,which uses the direction of two mark points and the relative position of two mark points to calculate the direction of parking slot.In order to verify the feasibility of the proposed method,the designed network was trained using the training set of the public dataset PS2.0,and the parking slot detection precision tested on the public dataset PS2.0 and the self-collected dataset PSS is 99.4%and 95.27%,the recall is 99.88%and 80.89%,the average error of the mark point position on PS2.0 is 0.84 pixel,and the error of the parking direction is 0.71 degree.The experimental results show that compared with the existing methods,the parking slot detection network proposed reduces the errors in the location of the mark point and the direction of the parking slot,and has a strong generalization ability on the PSS dataset.
作者
单凯强
桑海峰
Shan Kaiqiang;Sang Haifeng(School of Information Science and Engineering,Shenyang University of Technology,Shenyang 110870,China)
出处
《电子测量与仪器学报》
CSCD
北大核心
2022年第2期203-210,共8页
Journal of Electronic Measurement and Instrumentation
基金
国家自然科学基金(61773105)
辽宁省教育厅科研项目(LJGD2020006)资助
关键词
自动泊车
停车位检测
多尺度特征
标记点
深度学习
automatic parking
parking slot detection
multi-scale feature
mark point
deep learning
