基于关键点的复杂道路环境目标检测算法  

Complex Road Environment Target Detection Algorithm Based on Key Points

作　　者：董涛[1] 关潆 DONG Tao;GUAN Ying(School of Information Engineering,Liaodong University,Dandong 118003,China)

机构地区：[1]辽东学院信息工程学院,辽宁丹东118003

出　　处：《辽东学院学报（自然科学版）》2023年第1期57-64,共8页Journal of Eastern Liaoning University:Natural Science Edition

基　　金：辽宁省自然科学基金重点科技创新基地联合开放基金项目(2022-KF-12-13)。

摘　　要：针对复杂交通环境手工搭建神经网络人力成本高、小目标检测准确度低、使用锚框法参数多、算法实时性差等问题,提出一种基于关键点的复杂道路环境目标实时检测算法。首先,重构MBConv,改进EfficientNet主干特征提取网络,提高特征提取效率;其次,融入小尺度特征层,优化特征融合网络,提升复杂环境下小目标检测能力;最后,运用关键点预测法,完成检测目标分类及回归。在BDD100K数据集上的测试结果表明,设计算法的目标检测实时性较强,且对复杂环境中的小目标检测准确度较高。Aiming at the problems of high labor cost,low accuracy of small target detection,too many parameters of anchor frame method and poor real-time performance of algorithm in complex traffic environment,a real-time target detection algorithm based on key points in complex road environment was proposed.Firstly,MBConv was reconstructed,and EfficientNet backbone feature extraction network was improved so as to enhance feature extraction efficiency.Secondly,the small-scale feature layer was integrated to optimize the feature fusion network and improve the small target detection ability in complex environment.Finally,the key point prediction method was used to complete the detection target classification and regression.Tests on the BDD100K dataset showed that the method had strong real-time target detection and high accuracy for small target detection in complex environments.

关 键 词：特征提取 道路目标检测 特征融合 关键点预测 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]