基于二值网络的自动驾驶目标检测方法  

Automatic driving object detection method based on binary network

作　　者：吴岳敏[1] 孙圣鑫 王小龙 马彬[1] 程香平[4] WU Yue-min;SUN Sheng-xin;WANG Xiao-long;MA Bin;CHENG Xiang-ping(School of Electrical and Information Engineering,Hubei University of Automotive Technology,Shiyan 442002,China;Haoteng(Hubei)Intelligent Technology Co.,Ltd.,Shiyan 442000,China;School of Electronic Information Engineering,Beihang University,Beijing 100191,China;Institute of Applied Physics,Jiangxi Academy of Sciences,Nanchang 330029,China)

机构地区：[1]湖北汽车工业学院电气与信息工程学院,湖北十堰442002 [2]浩腾(湖北)智能科技有限公司,湖北十堰442000 [3]北京航空航天大学电子信息工程学院,北京100191 [4]江西省科学院应用物理研究所,江西南昌330029

出　　处：《陕西科技大学学报》2023年第2期176-183,共8页Journal of Shaanxi University of Science & Technology

基　　金：湖北省青年人才基金项目(Q20181804)。

摘　　要：针对现有全精度自动驾驶目标检测方法难以在车载计算资源受限平台实时部署等问题,提出了一种基于二值网络的目标检测方法.该方法通过重构残差网络单元和加宽每阶段通道数改进Faster R-CNN主干网络,以增强主干网络特征提取的能力.此外,该方法通过修改卷积核改进特征金字塔网络和区域提议网络,增强表征和预测能力.通过在两种常用目标检测数据集上进行的大量实验表明,该方法能够大幅度减小模型内存,提高检测速度,并取得与全精度模型相近的检测精度.该方法相比于其他先进的二值化目标检测算法,取得了最优秀的检测性能;相比全精度模型,平均参数量减少1.89倍,平均推理速度提高了6.10倍,而检测精度mAP在两数据集上分别仅下降0.2%和2.4%.Aiming at the problem that the existing full precision automatic driving detection methods are difficult to deploy in real time on the platform with limited computing resources,an object detection method based on binary network is proposed.This method improves the backbone network of Faster R-CNN by reconstructing residual network units and widening the number of channels in each stage,so as to enhance the capability of backbone network feature extraction.In addition,this method improves FPN and RPN by modifying the convolution kernel to enhance the representation and prediction capability.Extensive experiments on two target detection datasets show that this method can greatly reduce the model memory,improve the detection speed,and achieve the accuracy similar to the full accuracy model.Compared with other advanced binary target detection algorithms,this method has achieved the best detection performance;Compared with the full accuracy model,the average parameter quantity is reduced by 1.89 times,and the average reasoning speed is increased by 6.10 times,while mAP is only reduced by 0.2%and 2.4%on the two datasets,respectively.

关 键 词：自动驾驶 目标检测 二值化 特征金字塔网络 

分 类 号：U461.91[机械工程—车辆工程]