基于改进CenterNet的交通监控车辆检测算法  

Vehicle Detection Algorithm Based on Improved CenterNet in Traffic Surveillance

作　　者：成怡[1] 张宇[1] 李宝全 CHENG Yi;ZHANG Yu;Li Bao-quan(School of Control Science and Engineering,Tiangong University,Tianjin 300387,China)

机构地区：[1]天津工业大学控制科学与工程学院,天津300387

出　　处：《计算机仿真》2023年第1期131-136,212,共7页Computer Simulation

基　　金：国家自然科学基金(61973234);天津市自然科学基金(18JCYBJC88400);天津市自然科学基金(18JCYBJC88300)。

摘　　要：针对城市交通监控场景复杂的问题,提出一种基于CenterNet改进的车辆检测算法。选取ResNet101作为主干特征提取网络,引入自矫正卷积SCConv扩大网络感受野,改善模型结构。针对解码过程,采用深度可分离卷积并进行结构改进,增加网络宽度,并且施加SA注意力机制,抑制解码过程连续上采样产生的无用信息。改进后的网络在城市交通监控车辆检测中进行实验验证,实验结果表明,改进后的网络在车辆识别上mAP达到86.95%,较改进前提升了9.86%,适用于交通监控车辆检测任务。Aiming at the complexity of urban traffic surveillance scenes, an improved vehicle detection algorithm based on CenterNet network is proposed in this paper. ResNet101 is used as the backbone of CenterNet. Then Self-Calibrated Convolution is introduced to expand the receptive field of the network and improve the model structure. For the decoding process, the Depthwise Separable Convolution is selected and improve its structure to increase the network width. Then the SA attention mechanism is applied to suppress the useless information which is generated by continuous up-sampling in the decoding process. The improved network is experimentally verified in urban traffic surveillance vehicle detection. The experimental results show that the mAP of the improved network has reached 86.95% in vehicle identification, which is 9.86% higher than before and is suitable for the traffic surveillance vehicle detection task.

关 键 词：交通监控车辆检测 深度可分离卷积 注意力机制 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]