基于关键点和多帧图像特征融合的限高深度检测网络  

Height Limit Deep Detection Network Based on Key Points and Multi-Frame Image Feature Fusion

作　　者：刘路生 徐婕 崔峰 谢启伟 龙潜 LIU Lusheng;XU Jie;CUI Feng;XIE Qiwei;LONG Qian(School of Computer and Information Engineering,Hubei University,Wuhan 430062;Beijing Smarter Eye Technology Co.Ltd,Beijing 100023;Research Base for the Development of Modern Manufacturing Industry,Beijing University of Technology,Beijing 100124;College of Artificial Intelligence,Tianjin University of Science and Technology,Tianjin 300457)

机构地区：[1]湖北大学计算机与信息工程学院,武汉430062 [2]北京中科慧眼科技有限公司,北京100023 [3]北京工业大学现代制造业发展研究基地,北京100124 [4]天津科技大学人工智能学院,天津300457

出　　处：《系统科学与数学》2024年第7期1885-1901,共17页Journal of Systems Science and Mathematical Sciences

基　　金：国家重点研发计划(2020YFA0714201);国家重大研发计划(2018AAA0103103)资助课题。

摘　　要：路况检测是智能驾驶中的一项核心任务,其中包含限高检测任务.鉴于学术界中与限高检测相关的研究还不够成熟,文章对限高检测方法进行了研究,提出基于关键点和多帧图像特征融合的限高检测网络.将关键点思想引入限高检测任务,减少不必要的预测,提升检测效率;引入卷积门控循环单元(ConvGRU)对多帧图像进行建模,学习多帧图像之间的上下文关系,提升召回率,降低漏检率;提出空间细节特征(spatial particulars feature,SPF)模块,加强解码层的多尺度特征融合;引入坐标注意力机制,进一步关注目标检测区域,提升模型的查准率.实验结果表明:该网络不仅能够很好地完成限高检测任务,并且相比于BiSeNet、PINet、PSPNet等其他先进网络,能够更好地平衡查准率与召回率,拥有更高的F1值和较少的参数量;同时对于车道线检测任务,在精度与漏检率方面也表现优异,进一步证明了该网络的有效性.Road condition detection is a core task in intelligent driving,including height limit detection tasks.Considering that the research related to height limit detection in the academic community is not yet mature,we have conducted research on height limit detection methods and proposed a height limit detection network based on key points and multi-frame image feature fusion.By adopting key points in the height limit detection task,unnecessary predictions are reduced and detection efficiency is improved.By introducing a convolutional gated recurrent unit(ConvGRU)to model multiple images and learn the contextual relationship between multiple images,improve recall rate,and reduce missed detection rate.The spatial particulars feature(SPF)module is proposed,which strengthens the multi-scale feature fusion in the decoding layer.In order to improve the accuracy of the model,the coordinate attention mechanism is introduced,and the target detection area is further paid attention to.According to the experimental results,this network can not only complete the height limit detection task well,but also balance the precision and recall rate better,with higher F1 values and fewer parameters compared with other advanced networks such as BiSeNet,PINet,PSPNet,etc;At the same time,in the task of lane line detection,it also performs excellently in terms of accuracy and missed detection rate,further proving the effectiveness of the network.

关 键 词：深度学习 关键点 多帧图像 限高检测 智能驾驶 注意力机制 

分 类 号：U463.6[机械工程—车辆工程] TP391.41[交通运输工程—载运工具运用工程]