基于协作注意力和图神经网络的轻量化车位检测算法  

Lightweight Parking-Slot Detection Algorithm Based on Collaborative Attention and Graph Neural Network

作　　者：李琳辉[1,2] 袁世伟 连静[1,2] 顾汤鹏 Li Linhui;Yuan Shiwei;Lian Jing;Gu Tangpeng(School of Automotive Engineering,Dalian University of Technology,Dalian 116024;State Key Laboratory of Structural Analysis for Industrial Equipment,Dalian University of Technology,Dalian 116024)

机构地区：[1]大连理工大学,汽车工程学院,大连116024 [2]大连理工大学,工业装备结构分析国家重点实验室,大连116024

出　　处：《汽车技术》2023年第11期41-48,共8页Automobile Technology

基　　金：国家自然科学基金资助项目(61976039,52172382);中央高校基本科研业务费专项资金资助项目(DUT22JC09);大连市科技创新基金资助项目(2021JJ12GX015)。

摘　　要：为提高自动泊车过程车位检测的实时性和准确性,提出一种基于协作注意力和图神经网络的轻量化车位检测算法。首先,采用轻量化的网络结构,以改进的MobileNetV3作为特征提取网络,通过深度可分离卷积获得车位标记点的位置信息和特征信息,并将二者结合得到标记点的融合特征,然后构建图神经网络结构以增强车位标记点之间的内在联系,并结合协作注意力机制实现对多头注意力的整合,最后,通过公共车位数据集PS2.0对所提出的算法进行测试,结果表明,该算法的检测精度优于当前主流算法,平均每帧图像推理时间可缩短至10.1ms,具备良好的准确性和实时性。In order to improve the real-time and accuracy of parking slot detection in automatic parking,this paper proposed a lightweight parking-slot detection algorithm based on collaborative attention and graph neural network.Firstly,This algorithm used a lightweight network structure and the improved MobileNetV3 as the feature extraction network,obtained the location information and feature information of the parking-slot marker points through depthwise separable convolution,combined them to obtain the fused features of the marker points,then constructed a graph network structure to enhance the internal relationship of the parking-slot marker points,and combined the cooperative attention mechanism to integrate multiple attention.Finally,the algorithm was tested on the public parking-slot dataset PS2.0.The results indicate that the detection accuracy is better than the current mainstream algorithm,the average reasoning speed of each frame of image can reach 10.1 ms,with good accuracy and real-time performance.

关 键 词：车位检测 协作注意力 图神经网络 轻量化 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]