基于弹性融合的人群密度估算技术的应用

Application of pedestrian density estimation technology based on elastic fusion

人流密度估计被越来越多的机构所重视。由于人眼的估计速度和准确率无法满足应用需求,基于深度学习和卷积神经网络的相关理论,构建了一个卷积神经网络模型来实现人群密度检测。采用百度的paddle框架构建实验,使用YOLOv3目标检测+加权图融合方式,对不同人群数量的样本进行实验。对0~30人和50~540人分别进行了crowdnet和混合图融合模式模拟实验,通过调整学习率、采用Adam优化器、训练权重衰减率等参数调整,达到了93.68%的准确率,比近些年的crowdnet(87.84%)高出近6个百分点,达到了预期要求。

People density estimation has become the focus of more and more institutions.Because the speed and accuracy of human eye estimation can not meet the application needs,a convolution neural network model is constructed to realize the detection of population density after studying the relevant theories of deep learning and convolution neural network.Baidu′s paddle framework construction experiment is used,and the YOLOv3 target detection+weighted graph fusion method is used.The experiment studies on different populations,0~30 and 50~540 people carried out simulation experiments with crowdnet and mixed graph fusion mode,are carried out respectively.Through adjusting the learning rate,using Adam optimizer,training weight attenuation rate and other parameters repeatedly,the accuracy of 93.68%is finally achieved,It is nearly 6 percentage points higher than crowdnet(87.84%)in recent years,which meets the expected requirements.