Research On Head Detection Technology For Crowd Counting

As a typical computer vision task,the goal of crowd counting is to accurately and efficiently count the total number of people in the video or picture.At present,the task of crowd counting has been vigorously promoted and actively applied in many aspects of public security.More and more scholars begin to study it as a research topic.Based on the methods or means adopted,the crowd counting can be roughly divided into two types: one is the crowd counting based on regression,which mainly makes statistics of the number of people through the regression model created;the other is the crowd counting based on detection,which relies on mature target detection technology to make statistics of the number of people included in specific occasions.With the development of deep learning,the performance of crowd counting algorithm based on regression has made great progress by using convolutional neural network to learn the nonlinear mapping between image and density map.However,compared with the detection based method,density map has obvious limitations: it can only clarify the distribution characteristics of the crowd,and can't accurately locate the specific location of each head,which seriously restricts the promotion and application of this method in video monitoring,pedestrian recognition and other fields.In the task of crowd counting,head detection mainly faces a series of difficulties,such as low resolution of target,occlusion between objects and so on.In order to achieve accurate and efficient head detection in dense situations,this paper based on the two-stage object detection network model,and optimized and adjusted from the following two aspects:1)For the problem that the classification accuracy of small-scale head is low due to the low resolution and incomplete semantic information,this paper constructs a feature fusion network based on the feature pyramid network,which combines the semantic information of high-level features with the low-level features of high-resolution,so as to improve the network detection accuracy effectively.Then,by fusing the head distribution data extracted from the density map,the level of dense head detection is improved.2)The quality of density map will directly affect the effect of network detection of dense heads.In this paper,we try to adjust and optimize the density map prediction network by scientific and reasonable methods: by designing multi-channel convolution neural network,we can predict the density map in different branches according to the size of the target.Experimental results show that the performance of the density map prediction network designed in this paper is better than that of the existing regression counting algorithm.Finally,after feature selection and enhancement through softmax function and sigmoid function,the density map is integrated with the features extracted from the head detection network model,which makes the algorithm achieve excellent performance in the head detection task in the crowded visual scene and has good detection feasibility.