摘要
针对交通标志尺度变化大导致检测精度低的问题,本文提出一种改进CenterNet的交通标志检测算法。采用ResNeSt50作为主干特征提取网络,引入PSConv(Ploy-Scale Convolution)改进网络卷积层结构。设计多尺度感受野模块,对ASPP(Atrous Spatial Pyramid Pooling)选取合适的膨胀率,利用注意力机制优化模块输出,提升对不同尺度标志的检测能力。在解码网络设计特征增强模块,减少因连续上采样导致的特征丢失。针对CenterNet回归目标尺寸不准确的问题,使用GHM(Gradient Harmonizing Mechanism)对损失函数改进。经实验验证,改进后算法总体精度提升了9.45%,速度达到每秒91.09帧,适用于交通标志检测。
Aiming at the problem of low detection accuracy caused by the large scale change of traffic signs,this paper proposed a traffic sign detection algorithm based on improved CenterNet. ResNeSt50 was used as the backbone of improved CenterNet,and PSConv(Ploy-Scale Convolution)convolution was introduced to improve convolutional layers structure of the network. To improve the detection ability of different scale signs,this paper designed the multi-scale receptive field module which selected the appropriate expansion rate for ASPP(Atrous Spatial Pyramid Pooling)and used the attention mechanism to optimize the output of the module. The feature enhancement module was designed in the decoding network to reduce the feature loss caused by continuous up sampling. In order to inaccurate target size in CenterNet regression,GHM(Gradient Harmonizing Mechanism)was used to improve the loss function. Experimental results show that the overall accuracy of the improved algorithm is increased by 9. 45%,and the detection speed reaches91. 01 frames per second,which is suitable for traffic sign detection.
作者
成怡
张宇
李宝全
CHENG Yi;ZHANG Yu;LI Baoquan(School of Control Science and Engineering,Tiangong University,Tianjin 300387,China;Tianjin Key Laboratory of Intelligent Control of Electrical Equipment,Tianjin 300387,China)
出处
《信号处理》
CSCD
北大核心
2022年第3期511-518,共8页
Journal of Signal Processing
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)。
关键词
交通标志检测
多尺度感受野模块
特征增强模块
深度学习
traffic sign detection
multi-scale receptive field module
feature enhancement module
deep learning
