Research on single-stage multi-task YOLO-Parking algorithm based on obstacle and parking space detection

Journal Information

J. Hefei Univ. Tech. (Nat. Sci.), Jan.2024, Vol.47 No.1: 1-6, 61

DOI: 10.3969/j.issn.1003-5060.2024.01.001

Authors

ZHANG Bingli $ ^{1-2} $, WANG Yanhui $ ^{1-2} $, PAN Zehao $ ^{1-2} $, WANG Yixin $ ^{1-2} $, YANG Chenglei $ ^{1-2} $, WANG Xinyu $ ^{1-2} $

(1. School of Automobile and Traffic Engineering, Hefei University of Technology, Hefei 230009, China; 2. Anhui Intelligent Vehicle Engineering Laboratory, Hefei 230009, China)

Abstract and Keywords

Abstract: This paper proposes an end-to-end multi-task network model based on YOLOv4 for perception tasks in automatic parking systems. The model takes around view monitor (AVM) images as network inputs and extracts image feature information using convolutional networks. By utilizing YOLO and directional marking-point regression-parking slot (DMPR-PS) detection heads, it achieves parallel detection of parking spaces and obstacles. Validation on the PS 2.0 public dataset demonstrates that the proposed multi-task detection method can simultaneously detect parking spaces and obstacles, with an average precision of 89.72% for obstacle recognition and a precision rate of 93.53% for parking space recognition. The network achieves a detection rate of 34.0 frames per second, meeting the requirements of automatic parking perception tasks while improving detection efficiency. This research holds implications for the development of automatic parking perception technology.

Keywords: automatic parking; around view monitor (AVM); multi-task network; obstacle recognition; parking space recognition

作者信息

张炳力 $ ^{1,2} $，王焱辉 $ ^{1,2} $，潘泽昊 $ ^{1,2} $，王怿昕 $ ^{1,2} $，杨程磊 $ ^{1,2} $，王欣雨 $ ^{1,2} $

(1. 合肥工业大学汽车与交通工程学院，安徽合肥 230009；2. 安徽省智能汽车工程实验室，安徽合肥 230009)

摘要和关键词

摘要: 文章提出一种基于 YOLOv4 的端到端多任务网络模型用于自动泊车系统中的感知任务，以环视图像（around view monitor，AVM）作为网络输入，基于卷积网络提取图像特征信息，通过 YOLO 和 DMPR-PS（directional marking-point regression-parking slot）检测头实现停车位与障碍物并行检测。在 PS 2.0 公开数据集上进行验证的结果表明，所提出的多任务检测方法能够同时检测停车位和障碍物，障碍物识别平均精度均值达到 89.72%，车位识别查准率达到 93.53%，网络检测速率为 34.0 帧/s，在满足自动泊车感知任务需求的同时提升了系统的检测效率。该文研究成果对自动泊车感知技术的发展具有一定的意义。

关键词: 自动泊车；环视图像（AVM）；多任务网络；障碍物识别；停车位识别

Funding

长三角科技创新共同体联合攻关专项资助项目（2022CSJGG1501）；安徽省科技重大专项资助项目（202203a05020008）；安徽省发展和改革委员会2021新能源汽车产业创新发展资助项目（wfgcyh2021439）；中国声谷创新发展关键核心技术揭榜挂帅攻关资助项目（2108-340161-04-01-727575）和合肥市关键共性技术研发和重大科技成果工程化资助项目（2021CG003）

合肥工业大学学报(自然科学版)

Navigation