一种基于多尺度融合卷积神经网络的人群计数方法

A Crowd Counting Method Based on Multi-scale Fusion Convolutional Neural Network

人群间的相互遮挡和多变的空间尺度是基于单幅图像人群计数算法面临的主要挑战。近年来,基于深度学习的人群计数算法在该问题上取得了显著的成效,然而越来越深的网络结构给模型的训练和应用带来了困难。为了解决上述问题,提出了一种基于多尺度融合卷积神经网络(multi-scale fusion convolution neural network,MSF-CNN)的人群计数方法。方法采用三列不同大小卷积核的卷积神经网络来提取不同空间尺度的图像特征,同时在网络结构中引入融合层将提取到的特征进行融合并求取密度图,最后对密度图积分求和得到人群数量。在ShanghaiTech数据集及UCF_CC_50数据集上的实验结果表明,该方法能够适应复杂的场景,有效减少人群间相互遮挡和空间尺度的变化对计数结果的影响,同时模型易于训练,明显优于现有人群计数方法。

Crowd counting is a challenge task due to many factors such as mutual occlusion and scale variations.Although the counting algorithms based on deep learning have achieved remarkable successes in recent years on this issue,the deeper and deeper network structure brings difficulties to training and application of the model.To resolve these problems,a novel crowd counting method based on multi-scale fusion convolution neural network(MSF-CNN)was proposed.The MFCNN was composed of three column convolutional neural networks which were utilized to extract multi-scale image features,and these features were fused in a cooperative manner to obtain a crowd density map,and finally the population number was obtained by integrating the density map.The model was evaluated on ShanghaiTech and UCF_CC_50 datasets,and the experimental results demonstrate that the proposed method can adapt to various complex scenes and reduce the impact of occlusion and varied sizes to a large degree.Moreover,the network structure is relatively simple,and thus the model is easy to train,which is superior to some state-of-the-art methods.