Crowd Density Estimation Based On Multi-feature Fusion

With the development of science and technology,artificial intelligence has made major breakthroughs in multiple domains,such as big data analysis,computer vision,and semantic analysis,etc.As an important branch of artificial intelligence,Convolutional Neural Network(CNN)plays an important role in the field of computer vision.In recent years,CNN has made innovative breakthroughs time and time again in population density estimation research.Affected by spatial perspective,severe occlusion,light changes and other issues,population density estimation still faces a series of challenges in research.In the current research field of convolutional neural networks,the concept of multi-feature fusion exists in different forms,involving all aspects in the field.1.With the consideration of the following problems,such as the difficulty of detecting large and small targets in the dataset at the same time,the difficulty of processing deep semantic information and shallow detail information at the same time,and the shallow network architecture,a multi-scale and multi-column fusion network is proposed.In this model,feature maps of different sizes in vgg16 are fused to obtain feature maps with more informative content and more accurate and effective feedback of crowd density.The network framework is divided into three parts: feature extraction,feature fusion,and feature regression.In order to extract feature information more quickly and effectively,VGG16 is selected as the backbone network;in order to obtain more informative feature map,two feature layers of different sizes in the backbone network are fused;in order to detect targets of different sizes,multi-scale dilated convolution is incorporated to obtain different receptive fields.Through multi-scale and multi-column information fusion,the feature extraction ability of the network is enhanced.2.Aiming at the problem that the background area interferes with the final prediction result,a network with newly attention module,which integrates cross-level attention features,is proposed.The feature maps of different levels are merged,then the foreground and background areas are segmented to obtain the attention aggregation mechanism with robust perceptual ability.The network is divided into two parts.The first part selects the encoder-decoder structure of CSRNet to obtain feature information and performs density map regression.The second part uses the attention feature fusion module to obtain the density level map to reduce the interference of the background on the prediction results.Taking the feature maps of different levels in encoder as the input of the attention module leads to the high-aggregation attention perception of local and global features.Through the attention feature fusion module,the resistance of the network to the background is improved,and the accuracy of the prediction is higher.3.Summarize the characteristics of the methods in the first two chapters,combine the methods used in the two chapters,and propose a crowd density estimation method based on the attention mechanism under multi-level fusion.In order to overcome the above-mentioned problems at the same time,the first research method of crowd density estimation based on multi-scale and multi-channel fusion is combined with the second research method of crowd density estimation based on attention mechanism.Moreover,the density level map output by the attention feature fusion module is multiplied by the density map output by the multi-channel module to obtain more information while reducing background interference to obtain a more accurate density map.In order to verify the performance of the model,the test results of the Shanghai Tech dataset and UCF?CC?50 dataset are provided at the end of this article.The research results show that the density map of the network framework regression in this paper is more accurate,and the population density estimation results show higher indicators and have more superior robustness.