摘要
考虑到行人检测是视频监控领域的一项重要技术,其检测效果易受遮挡严重、光照不均等因素的影响,而人头检测是行人检测的重要研究内容,本文提出了一种基于混合卷积神经网络的人头检测方法。该方法将快速区域卷积神经网络(CNN)架构引入到局部模型的构建中,可以更好地获取图像的上下文信息,以得到更好的检测效果。通过全局模型预测头部的位置和尺度,利用成对模型确定待测目标间的成对关系。最后将局部、全局和成对模型融合成一个混合卷积神经网络框架,进行人头检测。研究结果表明,网络结构优化后的模型比多卷积神经网络方法在实时性显著提高52.3倍的同时,还可以将检测精度提高1.8%,计算复杂度和内存消耗也大大降低。
Based on the consideration that pedestrian detection is on important technique in video surveillance,and its detection effect is easy to be affected by serious occlusion,uneven illumination,etc.,while human head detection is an important part of pedestrian detection,a human head detection method based on hybrid convolution neural networks（ CNNs） is proposed. The method introduces fast regional convolutional neural network architecture into the construction of the local model for obtaining context image information for detecting person better. The global model is built to predict the position and scale of the head. The pairwise model is used to get the pairwise from objectives. At last,the local,global and pairwise models are fused into a federated CNN framework in order to detect head. Compared with the context CNNs,the research achievement shows that the hybrid CNNs can improve the real-time and average precision 52. 3 times and 1. 8 percent individually. Computational complexity and memory consumption can be reduced significantly.
作者
吉训生
吴凡
Ji Xunsheng;Wu Fan(School of Internet of Things Engineering,Jiangnan University,Wuxi 214122)
出处
《高技术通讯》
EI
CAS
北大核心
2018年第4期313-319,共7页
Chinese High Technology Letters
基金
国家自然科学基金(61771223)
江苏省前瞻性联合研究(BY2016022-28)资助项目
关键词
图像处理
行人检测
人头检测
上下文
卷积神经网络(CNN)
迁移学习
image processing
pedestrian detection
head detection
context
convolution neural network（CNN）
transfer learning
