光电技术应用, 2019, 34 (6): 40, 网络出版: 2019-12-08
基于多信息卷积神经网络的人群密度估计
Crowd Density Estimation Based on Multi-information via Convolutional Neural Network
计算机视觉 人群密度估计 卷积神经网络 密度图 computer vision crowd density estimation convolutional neural network density map
摘要
针对现有人群密度估计方法在实际应用中容易受到环境干扰、适应性不强等限制, 提出一种基于多信息卷积神经网络的人群密度估计方法。首先, 根据样本数据的特点生成数据密度图标注与数据增强; 然后, 为适应不同真实场景的巨大差异, 提取图像的色调、色饱和度、灰度(HSG)信息作为训练数据的输入, 并利用共享卷积层、结合两个子网络不同卷积深度的特征构建网络模型; 最后, 对网络输出的密度图进行积分, 得到相应的人数。与主流方法对比, 在Shanghaitech数据集上进行的相关实验证明了所提方法的良好性能。
Abstract
In view of the limitations of existing crowd density estimation methods in practical application, such as environmental interference and weak adaptability, a crowd density estimation method based on multi-information via convolution neural network is proposed. Firstly, data density annotation and data augmentation are implemented according to the characteristics of samples. And then, in order to adapt to different real scenarios with large variations, the proposed method is used to extract the hue, saturation and gray level (HSG) information of images as the input of training data, and the shared convolutional layers are utilized, combination of features from two subnetworks with different convolution depth to construct the network model. Finally, the output density map is integrated to get the corresponding number of people. Compared with the typical methods, experiments on Shanhaitech datasets demonstrates the great performance of the proposed method.
赵威, 吴晓红, 刘文璨, 何小海, 卿粼波. 基于多信息卷积神经网络的人群密度估计[J]. 光电技术应用, 2019, 34(6): 40. ZHAO Wei, WU Xiao-hong, LIU Wen-can, HE Xiao-hai, QING Lin-bo. Crowd Density Estimation Based on Multi-information via Convolutional Neural Network[J]. Electro-Optic Technology Application, 2019, 34(6): 40.