关键词: 图像, 密集人群, 人群计数, 特征提取, 神经网络, 单列式神经网络, 改进残差结构

Abstract:

In order to avoid crowd trampling, it is very important to accurately obtain information on the number of crowds from surveillance images. Early crowd counting studies used a feature engineering approach, in which human-designed feature extraction algorithms were used to obtain features that represented the number of people to be counted. However, the counting accuracy of such methods is not sufficient to meet the practical requirements when facing heavily occluded counting scenes with large changes in scene scale. In recent years, with the development of neural network, breakthroughs have been made in image classifications, object detections, and other fields. Neural network methods have also advanced the accuracy and robustness of dense crowd counting. In view of the difficulty of counting dense crowds, small crowd targets, and large changes in scene scale, this paper proposes a new neural network structure named: VGG-ResNeXt. The features extracted by VGG-16 are used as general-purpose visual description features. ResNet has more hidden layers, more activation functions and has more powerful feature extraction capabilities to extract more features from crowd images. Improved residual structure ResNeXt expands on the base of ResNet to further enhance feature extraction capabilities while maintaining the same computing power requirements and number of parameters. Therefore, in this paper, the first 10 layers of VGG-16 are used as the coarse-grained feature extractor, and the improved residual neural network ResNeXt is used as the fine-grained feature extractor. With the improved residual neural network feature of "multi-channel, co-activation", the single-column crowd counting neural network obtains the advantages of the multicolumn crowd counting network (i.e., extracting more features from dense crowd images with small targets and multiple scales), while avoiding the disadvantages of the multicolumn crowd counting network, such as the difficulty of training and structural redundancy. The experimental results show that our model achieves the highest accuracy in the UCF-CC-50 dataset with a very large number of people per image, the ShangHaiTech PartB dataset with a sparse crowd, and the UCF-QNRF dataset with the largest number of images currently included. Our model outperforms other models in the same period by 7.5%, 18.8%, and 2.4%, respectively, in MAE in the above three datasets, demonstrating the effectiveness of the model in improving counting accuracy in dense crowds. The results of this research can effectively help city management, relieve the pressure on public security, and protect people's lives and property.

Key words: images, dense crowd, crowd counting, feature extraction, neural networks, single column-based CNN, improved ResNet