基于ResNet-MLP模型的车辆目标检测算法  

Vehicle target detection algorithm based on ResNet-MLP modeling

作　　者：王可栋 曲含章 马敏 杨子奕[3] 康爱平 WANG Kedong;QU Hanzhang;MA Min;YANG Ziyi;KANG Aiping(Intelligent Manufacturing Institute,Qingdao Huanghai University,Qingdao 266427,China;School of Mechanical and Automotive Engineering,Qingdao University of Technology,Qingdao 266525,China;School of Civil Engineering,Qingdao University of Technology,Qingdao 266525,China;Chang’an Dublin International College of Transportation,Chang’an University,Xi’an 710064,China;Qingdao Hicorp Heavy Industry Science&Technology Co.Ltd.,Qingdao 266400,China;College of Energy and Mining Engineering,Shandong University of Science and Technology,Qingdao 266590,China;College of Business and Economics,Lanzhou Modern Vocational College,Lanzhou 730300,China)

机构地区：[1]青岛黄海学院智能制造学院,青岛266427 [2]青岛理工大学机械与汽车工程学院,青岛266525 [3]青岛理工大学土木工程学院,青岛266525 [4]长安大学长安都柏林国际交通学院,西安710064 [5]青岛环球重工科技有限公司,青岛266400 [6]山东科技大学能源与矿业工程学院,青岛266590 [7]兰州现代职业学院财经商贸学院,兰州730300

出　　处：《青岛理工大学学报》2024年第3期142-150,共9页Journal of Qingdao University of Technology

基　　金：国家自然科学基金资助项目(52272311)。

摘　　要：为提高车辆目标检测精度,针对传统视觉传感器光照敏感性、空间感知性差等缺点,采用激光雷达传感器,提出一种基于ResNet-MLP二阶段模型的车辆目标检测算法。该算法对点云鸟瞰图的映射方式进行改进,使其保留点云高度特征,并通过改进后的ResNet进行点云特征的提取,最后使用并行多层感知机网络对车辆目标分类和位置回归。采用KITTI的3D Object数据集进行验证,通过与PointNet++和VoxelNet方法进行对比实验,结果发现,交并比(IOU)较高时3种方法的检测精度均有所下降,但相对于其他2种算法,本算法检测精度更高,运行速度更快,可为未来自动驾驶车辆的实时感知方面提供技术支撑。Aiming at the disadvantages of low sensitivity and poor spatial perception of traditional visual sensors,a vehicle target detection algorithm based on ResNet-MLP two-stage model is proposed to enhance the accuracy of vehicle target detection by using LiDAR sensor.The algorithm improves the mapping of point cloud bird’s-eye view so that it retains the point cloud height features,and the point cloud features are extracted by the improved ResNet.Finally,a parallel multilayer perceptron network is used to classify vehicle targets and regress the position.KITTI’s 3D Object dataset is used for validation,and through comparison experiments with PointNet++and VoxelNet methods,it is found that the detection accuracy of the three methods decreases when the IOU is higher,but compared to the other two algorithms,the proposed algorithm has better detection accuracy and higher running speed,which can provide technical support for the real-time perception of self-driving vehicles in the future.

关 键 词：车辆目标检测 激光雷达 点云图像 鸟瞰图映射 

分 类 号：U491.1[交通运输工程—交通运输规划与管理]