基于语义分割的车位检测算法研究  

Research on parking space detection algorithm based on semantic segmentation

作　　者：李伟东[1] 李冰[1] 朱旭浩 李乐[1] LI Weidong;LI Bing;ZHU Xuhao;LI Le(School of Automobile Engineering,Dalian University of Technology,Dalian 116024,China)

机构地区：[1]大连理工大学汽车工程系,辽宁大连116024

出　　处：《大连理工大学学报》2024年第1期96-103,共8页Journal of Dalian University of Technology

基　　金：辽宁省重点研发计划资助项目(2020JH2/10100028)。

摘　　要：作为自动泊车系统中至关重要的一环,车位检测算法的精度直接决定自动泊车系统的好坏.目前,基于语义分割的车位检测算法主要有两个问题:一是分割网络参数量较大,难以满足移动端部署;二是后处理提取算法复杂,难以满足实时检测要求.针对这两个问题,设计一种通过检测车位线来获取停车位的车位检测算法.采用深度可分离卷积和非对称卷积相结合的方式设计车位线分割网络UFAC-Net,并提出一种更为简洁的车位线提取算法.实验结果表明:UFAC-Net模型(UFAC-Net2)分割的平均像素精度为83.07%,平均交并比为73.05%,模型参数量为3.1 MB,达到目前PSV datasets上最好的分割精度;车位检测算法可检测复杂情况下的平行、垂直、倾斜3种类型的车位,在自定义测试集中精准率为99.23%,召回率为99.12%,单张图像检测时间为32.2 ms,具有良好的检测性能.As an important part of automatic parking system,the accuracy of parking space detection algorithm directly determines the quality of automatic parking system.At present,the parking space detection algorithm based on semantic segmentation has two main problems:first,the number of segmentation network parameters is too large to meet the requirements of mobile terminal deployment;second,the post-processing extraction algorithm is too complicated to meet the requirements of real-time detection.To solve these two problems,a parking space detection algorithm is designed to obtain parking space by detecting parking lines.By combining the depth separable convolution and asymmetric convolution,a network UFAC-Net for parking line segmentation is designed,and a more concise algorithm for parking line extraction is proposed.The experimental results show that the mean pixel accuracy of UFAC-Net model(UFAC-Net2)segmentation is 83.07%,the mean intersection over union is 73.05%,and the number of model parameters is 3.1 MB,which achieves the best segmentation accuracy of the PSV datasets.Parking space detection algorithm can extract parallel,vertical and inclined parking spaces in complex situations.The accuracy rate of parking space detection algorithm in the custom test set is 99.23%,the recall rate is 99.12%,and the detection time of a single image is 32.2 ms,which show that the proposed algorithm has good detection performance.

关 键 词：车位检测 语义分割 深度可分离卷积 非对称卷积 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]