摘要
道路目标检测在智慧城市建设中扮演着重要角色,而Faster-RCNN是目前主流的目标检测网络结构算法.本文在Faster-RCNN卷积神经网络结构基础上增加了特征金字塔网络层,并采用关注损失函数替代了原有的交叉熵损失函数.其中增加的特征金字塔特征融合层可以提取到检测图片中更具鲁棒性和一般性的前背景特征,而通过关注损失函数则能起到缓解检测图片中的正负样本不均的情况.最后,在公开数据集KITTI上实验证实,改进的目标检测算法能实现提高原有的Faster-RCNN目标检测准确率.
Road object detection plays a vital role in intelligent city construction;and Faster-RCNN is one of the main stream object detection algorithms.The paper proposed a concatenated feature pyramid network layer on the original Faster-RCNN feature extraction layer and substituted the cross entropy loss function with the focal loss function.The concatenated feature pyramid network could extract the enriched feature maps that were more robust and generalized to diverse situations.The adopted focal loss function could alleviate the sample inhomogeneity in the detected picture.The algorithm was verified by using open KITTI datasets.The result shows that the updated Faster-RCNN algorithm can improve detection precision.
作者
陈飞
章东平
CHEN Fei;ZHANG Dongping(College of Information Engineering,China Jiliang University,Hangzhou 310018,China)
出处
《中国计量大学学报》
2018年第4期393-397,共5页
Journal of China University of Metrology
基金
浙江省自然科学基金项目(No.LY15F020021)
浙江省公益技术应用研究计划项目(No.2016C31079)
