摘要
车辆在自适应巡航过程中,巡航控制系统的信息输入主要由毫米波雷达和视觉传感器来提供,这就要求车辆处于高速行驶或环境信息复杂的路况下具备良好的目标检测与识别能力。为了提升车辆巡航控制系统中基于视觉传感器的车辆目标检测结果,在FPN模型基础上,提出一种基于深度监督特征融合的ADAS车辆目标检测算法,利用多级特征跳跃融合策略,能有效弥补网络编码阶段多次下采样带来的信息损失,通过深度监督策略,提高了网络的辨别力和稳健性。实验结果表明,相较于几种对比方法,基于深度监督特征融合的车辆目标检测算法在车辆图像数据集上实现了较高的检测精度,体现出了该方法的有效性和优越性。
During the adaptive cruise process of the vehicle,the information input of the cruise control system is mainly provided by millimeter wave radar and visual sensors,which requires the vehicle to have good target detection and recognition capabilities in high-speed driving or road conditions with complex environmental information.In order to improve the vehicle target detection results based on visual sensors in the vehicle cruise control system,based on the FPN models,a vehicle target detection algorithm using depth monitoring and supervision feature fusion is proposed.By using a multi-level feature jump fusion strategy,this method can effectively compensate for the information loss caused by multiple down-sampling in the network coding stage.The experimental results show that,compared with several comparison methods,the vehicle target detection algorithm based on deep supervision feature fusion achieves higher detection accuracy on the vehicle image data set,which reflects the effectiveness and superiority of the method.
作者
雷嘉豪
李江
刘波
LEI Jiahao;LI Jiang;LIU Bo(School of Automotive Engineering,Shaanxi Polytechnic Institute(Xianyang,Shaanxi,712000,China))
出处
《小型内燃机与车辆技术》
2022年第2期58-63,共6页
Small Internal Combustion Engine and Vehicle Technique
基金
陕西工业职业技术学院院级科研计划项目(2020YKYB-026)。
关键词
深度学习
高级辅助驾驶(ADAS)
实例分割
特征融合
深度监督
Deep learning
Advanced driving assistance system(ADAS)
Instance segmentation
Feature fusion
Deep supervision
