轻量化车位识别网络与流畅泊车路径规划设计  

Lightweight Recognition Network for Parking Space and Smooth Parking Path Planning Algorithm Design

作　　者：王玉龙 李景俊[1] 翁茂楠 黄辉 WANG Yu-long;LI Jing-jun;WENG Mao-nan;HUANG Hui(Auto Engineering Research Institute of Guangzhou Automobile Group,Guangzhou 510641,China;State Key Laboratory of Advanced Design and Manufacturing for Vehicle Body,Changsha 410082,China)

机构地区：[1]广州汽车集团股份有限公司汽车工程研究院,广州510641 [2]湖南大学汽车车身先进设计制造国家重点实验室,长沙410082

出　　处：《汽车科技》2023年第2期2-7,共6页Auto Sci-Tech

基　　金：湖南大学汽车车身先进设计制造国家重点实验室开放基金(31825011)。

摘　　要：深度学习在自动驾驶开发中得到广泛应用,而在低算力嵌入式平台上部署高算力需求的车位识别网络及复杂的路径规划算法成为行业挑战,因此本文首先设计了轻量化车位识别网络结构,对不同场景下的车位关键点进行检测,其次基于深度搜索设计了快速泊车路径规划算法,然后基于OpenVX的框架将模型和算法部署在多核异构平台上,最终在实车平台上进行自动泊车功能验证。试验结果表明车位检出率大于98%,泊入成功率大于96%,泊车系统运行时间小于40ms,满足实时性要求。Deep learning is widely used in the development of autonomous driving,however the deployment of parking space recognition network and complex path planning algorithm with high computing power requirements on low computational embedded platform has become an industry challenge.Therefore,this paper first designs a lightweight parking space recognition network structure to detect the key points of parking space in different scenarios,then designs a fast parking path planning algorithm based on deepsearch algorithm,and then deploys the model and algorithm on a multi-core heterogeneous platform based on the OpenVX framework,and finally performs automatic parking function verification on the real vehicle platform.The test results show that the detection rate of parking space is more than 98%,and the parking success rate is more than 96%,and the parking system running time is less than 40ms,which meets the real-time requirements.

关 键 词：车位识别 深度学习 自动泊车 路径规划 

分 类 号：U463.3[机械工程—车辆工程]