湖南大学学报(自然科学版)2024,Vol.51Issue(2) :177-186.DOI:10.16339/j.cnki.hdxbzkb.2024156

基于全景环视的端到端停车位检测方法

End-to-end Parking Slot Detection Method Based on Panoramic Surround View

秦晓辉 殷周平 张素英 黄圣杰 张润邦 芦涛 刘硕 胡满江
湖南大学学报(自然科学版)2024,Vol.51Issue(2) :177-186.DOI:10.16339/j.cnki.hdxbzkb.2024156

基于全景环视的端到端停车位检测方法

End-to-end Parking Slot Detection Method Based on Panoramic Surround View

秦晓辉 1殷周平 2张素英 3黄圣杰 2张润邦 2芦涛 2刘硕 2胡满江1
扫码查看

作者信息

  • 1. 湖南大学 机械与运载工程学院,湖南 长沙 410082;湖南大学 无锡智能控制研究院,江苏 无锡 214072
  • 2. 湖南大学 机械与运载工程学院,湖南 长沙 410082
  • 3. 潍柴动力股份有限公司 发动机研究院,山东 潍坊 261061
  • 折叠

摘要

现有停车位检测方案大多将目标检测方案和人工设计的后处理模块进行简单结合,各阶段提取的特征存在大量冗余信息.并且,人工设计的后处理模块通常适应面窄,计算量大,最终导致停车位检测效果难以实用.针对这些问题,本文引入全景视觉,结合现有算法的优点与环视图像的特点,设计端到端的无锚框停车位检测算法.该算法对停车位进行进入线朝向建模,而非单独考虑两个入口点,省去了停车位入口点匹配和朝向判断流程,最终实现完全一体化的停车位位置、朝向和占用情况检测.考虑到实用性,在网络结构设计上进行了速度和精度的平衡、正负样本均衡,以及无后处理等多方面优化.最终,在ps2.0数据集上,本文提出的AFPSD模型以88.7的FPS(每秒帧数)达到68.7%的AP,相较VPS-Net和DMPR-PS方案精度分别提升1.2%和2.1%.由此可知,本文设计的一阶段端到端方案可以代替三阶段方案,在环视图像上实现停车位的稳定检测.

Abstract

Most of the existing parking spot detection solutions simply combine the target detection scheme and with manually designed post-processing modules,and there is a large amount of redundant information in the features extracted at each stage.Moreover,the manually designed post-processing modules are usually narrowly adapted and computationally intensive,which ultimately makes the parking space detection effect difficult to be practical.To address these problems,this paper introduces panoramic vision and combines the advantages of existing algorithms with the characteristics of surround-view images to design an end-to-end anchorless frame parking spot detection algorithm.The algorithm models the entry line orientation of parking spaces instead of considering two entry points separately,eliminating the process of parking space entry point matching and orientation judgment,and finally realizing realizes fully integrated parking space location,orientation,and occupancy detection.Considering the practicality,the network structure design is optimized in many aspects,such as the balance of speed and accuracy,positive and negative sample balance,and no post-processing.Finally,on the ps2.0 dataset,the AFPSD model proposed in this paper achieves 68.7%AP with a FPS(Frames Per Secend)of 88.7,which is 1.2%and 2.1%higher accuracy compared to the VPS-Net and DMPR-PS schemes,respectively.It can be seen that the one-stage end-to-end scheme designed in this paper can replace the three-stage scheme to achieve stable detection of parking slots on the surround-view image.

关键词

自主泊车/目标检测/停车位检测/环视图像

Key words

autonomous parking/object detection/parking slot detection/surround-view image

引用本文复制引用

基金项目

长沙市自然科学基金资助项目(kq2202162)

国家自然科学基金资助项目(52172384)

湖南省湖湘青年科技创新人才(2021RC3048)

汽车车身先进设计制造国家重点实验室自主研究课题(61775006)

出版年

2024
湖南大学学报(自然科学版)
湖南大学

湖南大学学报(自然科学版)

CSTPCD北大核心
影响因子:0.651
ISSN:1674-2974
参考文献量21
段落导航相关论文