基于改进深度卷积神经网络的车辆检测方法研究  

Research on Vehicle Detection Method Based on Improved Deep Convolutional Neural Network

作　　者：严友 YAN You(Quzhou College of Technology,Quzhou 324000,China)

机构地区：[1]衢州职业技术学院,浙江衢州324000

出　　处：《汽车电器》2022年第1期5-7,10,共4页Auto Electric Parts

基　　金：基于智能行人检测汽车辅助安全驾驶技术研究,衢州市科技计划指导性项目(2018006)。

摘　　要：基于视觉的车辆检测作为辅助驾驶系统的输入,对智能车辆预警和决策起着重要的作用。针对目前传统深度卷积神经网络在基础网络设计和物体检测网络构建的不足,提出一种对经典的深度残差网络进行改进方法,提出带局部连接的残差单元,并以此构建带局部连接的残差网络;同时,提出基于共享参数的多分支网络和双金字塔语义传递网络形式,提升不同语义级别特征融合前的语义级别,以及实现深度融合不同分辨率特征图的语义。经过测试,车辆的检测准确率最高达到95.3%,且具备较高的实时性和环境适应性。Vision-based vehicle detection,as the input of the driving assistance system,plays an important role in the early warning and decision-making of intelligent vehicles.Aiming at the current shortcomings of traditional deep convolutional neural networks in basic network design and object detection network construction,a method to improve the classic deep residual network is proposed,and a residual unit with local connections is proposed,and a local connection is constructed based on this.At the same time,a multi-branch network and a double-pyramid semantic transfer network form based on shared parameters are proposed to improve the semantic level before the feature fusion of different semantic levels,and to achieve the deep integration of the semantics of feature maps of different resolutions.After testing,the vehicle's detection accuracy rate is up to 95.3%,and it has high real-time and environmental adaptability.

关 键 词：智能汽车 深度卷积神经网络 车辆检测 算法验证 

分 类 号：U463.6[机械工程—车辆工程]