Deep Learning-Based People Counting In Video Scenes

People counting in video scenes is an important application of Computer Vision.It is of great significance in public security,site planning,crowd guidance,group management and so on.However,because of the complexity and diversity of video scenes,the traditional approaches may not be accurate and robust enough.Deep learning is of powerful feature extraction ability,non-linear expression ability and generalization ability.Therefore,starting with deep learning,this paper studies and designs the people counting algorithm in video scenes based on deep learning.This paper is aimed at two kinds of surveillance video scenes:the first one is video with top-down view,which often appears in building entrance or elevator;and the second one is video with inclined view,which often appears at roadside or inside public places such as shopping malls.According to the characteristics of these two kinds of video,this paper proposes two kinds of deep learning-based people counting algorithms in video scenes,which are respectively based on pedestrian detection and crowd density map.As for the people counting algorithm based on pedestrian detection,in this video scene with top-down view,the coverage of video is small and there are fewer pedestrians while most pedestrians are clearly visible and makes for easy detection.But,as the main detection targets,the head and shoulders of the pedestrian is relatively small,and the pedestrians' semantically affiliated attributes(i.e.the pedestrians' carry-on items such as backpacks and hats)easily lead to wrong or missed detection.Although deep learning has made great achievements in the field of target detection,and even in multiple-target detection,there is still no deep learning algorithm that's robust to small targets.To solve these problems,this paper proposes a method for pedestrian detection based on deep learning joint with semantic attributes:Firstly,we improve the ability of the neural network to detect small targets.Secondly,the pedestrian's semantic attributes which can easily cause false or missed detection,are used as auxiliary detection targets.Combining the pedestrian with its semantic attributes increases the reliability of pedestrian target detection.The experimental results show that the people counting method based on pedestrian detection is fast and of high precision,and practical value.As for the people counting algorithm based on crowd density map,in this video scene with inclined view,the coverage of video and the size of the crowd is very large,the pedestrians in different locations in the video are very different,and the occlusion and overlap of pedestrians are very serious.In that case,the crowd distribution is important information because the group incidents often happen in crowded places.Therefore,this paper designs a convolutional neural network,mapping the video images to the crowd density map,and then calculating the number of people in the image with the crowd density map.In this paper,a large number of complex features are extracted from video images by a deeper and wider network structure.The network combines features of high and low dimension with residual connection,and extracts the features of different scales by using multi-scale convolution kernel to deal with the pedestrian scale change.The experimental results show that the convolutional neural network designed for this kind of scenes is of high accuracy.In addition,this paper uses the context relationship between video frames to further improve the counting results of these two algorithms.Since the number of people in the adjacent frames is similar,we use this property to further improve the accuracy of the counting results.The experimental results also show that the inter-frame result correction is effective.