基于SSD-MobilenetV3模型的车辆检测  

Vehicle detection based on SSD-MobilenetV3 model

作　　者：廖慕钦 周永军 汤小红[1] 蒋淑霞[1] 李宇琼 LIAO Muqin;ZHOU Yongjun;TANG Xiaohong;JIANG Shuxia;LI Yuqiong(School of Mechanical and Electrical Engineering,Central South University of Forestry and Technology,Changsha 410000,China)

机构地区：[1]中南林业科技大学机电工程学院,湖南长沙410000

出　　处：《传感器与微系统》2022年第6期142-145,共4页Transducer and Microsystem Technologies

基　　金：长沙市科技计划资助项目(KQ1701102)。

摘　　要：针对自动驾驶平台车辆检测问题,提出一种结合迁移学习的卷积神经网络(CNN)模型SSD-MobilenetV3。网络结合SSD检测速度较快与MobilenetV3占用内存小的优点,将SSD模型的基础网络替换成MobilenetV3。首先,结合迁移学习的方法,在COCO数据集上对网络进行预训练,再使用自建融合车辆数据集对预训练模型全连接层进行重新训练,可在短时间训练下得到收敛,并有较好的准确率。实验结果表明:相比原SSD模型,检测准确率达到85.6%,提高了3.1%;参数量减为16.9 Mbyte,减少了83.1%。模型在准确率小幅上升的同时,大幅度减少占用内存,更适用于自动驾驶平台。A convolutional neural network model,SSD-MobilenetV3,which combines transfer learing SSD detection speed with MobilenetV3,is proposed to solve the problem of vehicle detection on autonomous driving platform.The network replaces the basic network of SSD model with MobilenetV3 with the advantages of faster SSD detection speed and less memory occupied by MobilenetV3.Firstly,combined with the transfer learning method,the network is pretrained on the COCO dataset.And then,the self-built fusion vehicle dataset is used to retrain the full connection layer of the pretraining model,which can be converged in a short time with good accuracy.The experimental results show that compared with the original SSD model,the detection accuracy increased by 3.1%,reaches 85.6%.The number of parameters is reduces by 83.1%to 16.9 Mbyte.While the accuracy of the model increases slightly,the memory occupied by the model is greatly reduced,which is more suitable for automatic driving platform.

关 键 词：自动驾驶 车辆检测 迁移学习 SSD-MobilenetV3模型 

分 类 号：TP389.1[自动化与计算机技术—计算机系统结构]