基于改进型CenterNet的车辆检测应用  

Application of vehicle detection based on improved CenterNet

作　　者：梁礼明 熊文 钱艳群 蓝智敏 LIANG Liming;XIONG Wen;QIAN Yanqun;LAN Zhimin(School of Electrical Engineering and Automation,Jiangxi University of Science and Technology,Ganzhou 341000,China)

机构地区：[1]江西理工大学电气工程与自动化学院,江西赣州341000

出　　处：《现代电子技术》2021年第11期141-145,共5页Modern Electronics Technique

基　　金：国家自然科学基金(51365017);国家自然科学基金(61463018);江西省自然科学基金面上项目(20192BAB205084);江西省教育厅科学技术研究重点项目(GJJ170491)。

摘　　要：CenterNet算法在车辆检测领域中表现优异,具有检测精度高和速度快的特点,但其也具有明显的缺点,由于网络采用复杂的Hourglass-104结构导致检测速度达不到实时性要求,同时检测过程中也有误检和漏检的现象发生。针对这些问题,提出一种基于改进型CenterNet的车辆检测算法。该方法首先对网络结构进行精简,将2个堆叠的Hourglass网络中的下采样和上采样次数减少为3次,然后用空洞卷积替换传统卷积以增大网络的感受野,捕捉多尺度上下文信息,最后通过在残差单元中增加一条支路以实现检测精度的提升。在KITTI数据集上进行仿真实验,检测精度和每张图片的检测时间分别为58.9%和105 ms,总体性能优于现有算法。The CenterNet algorithm performs well in the field of vehicle detection.It has the characteristics of high detection precision and detection speed.However,it has obvious disadvantages,too.Due to complex Hourglass-104 structure adopted in the network,the detection speed fails to meet the real-time requirement,and the false detection and missing detection also occur in the detection process.In view of the above,an improved vehicle detection algorithm based on CenterNet is proposed.In the method,the network structure is simplified,both the downsample times and upsample times in the two stacked Hourglass network are reduced to 3 times,and then the traditional convolution is replaced with the dilated convolution to increase network receptive field and capture multi-scale context information.Finally,a branch is added to the residual unit to improve the detection precision.Simulation experiments were performed with KITTI data sets.The detection precision and detection duration per picture of the proposed method is 58.9%and 105 ms,which is superior to those of the existing algorithms.

关 键 词：改进型CenterNet 车辆检测 HOURGLASS 残差单元 空洞卷积 锚点框 网络精简 消融实验 

分 类 号：TN911.73-34[电子电信—通信与信息系统] TN711[电子电信—信息与通信工程]