摘要
针对现实生活中存在人与人之间相互重叠交叉遮挡,由此产生对行人检测技术中检测速度慢、检测准确率低以及鲁棒性较差等问题。实验基于YOLOv3网络架构,为减少网络传递过程逐层丢失信息,借鉴残差密集网络的思想,提出一种改进YOLOv3算法,实现网络多层特征复用及融合,并以扩增数据集以及多尺度策略等方法训练网络。实验结果表明:与目前主流目标检测方法相比,该方法提高了有遮挡行人检测准确率与召回率。
Aiming at the problems of slow detection,low detection accuracy and poor robustness in pedestrian detection technology generated by overlapping occlusion between people in real life,based on the YOLOv3 network architecture,propose an improved YOLOv3 algorithm to realize network multi-layer feature multiplexing and fusion,and to amplify datasets and methods such as scale strategy train the network.The experimental results show that compared with the current mainstream target detection methods,this method improves the detection accuracy and recall rate of occluded pedestrians.
作者
梁礼明
邓广宏
卢明建
吴健
LIANG Liming;DENG Guanghong;LU Mingjian;WU Jian(School of Electrical Engineering and Automation,Jiangxi University of Science and Technology,Ganzhou 341000,China)
出处
《传感器与微系统》
CSCD
2020年第5期148-151,共4页
Transducer and Microsystem Technologies
基金
国家自然科学基金资助项目(51365017,61463018)
江西省自然科学基金资助项目(20132BAB203020)
江西省教育厅科学技术研究重点资助项目(GJJ170491)。