Research On Crowd Counting In Dense Scene

Crowd Counting is an important research branch in the field of computer vision.Its purpose is to accurately predict the number,distribution,and overall density of pedestrians in real scenes.Because it can provide accurate and detailed crowd status data in real scenes,crowd counting technology is widely used in dense scenes,such as parades,tourist attractions,shopping malls,etc.By accurately estimating the number of crowds,especially the crowd status information in dense scenes,crowded or abnormal behaviors can be detected in time and alerted,so that measures can be taken to divert and avoid safety accidents.Therefore,the study of crowd counting has important application value and practical significance.However,in the actual scene application,there are many problems,such as complex background interference,rapid change of pedestrian scale,chaotic crowd distribution,and so on,which bring great challenges to the task of crowd counting.Therefore,to overcome the challenges posed by the above problems,this paper studies the network structure,loss function,and other aspects of the algorithm to overcome the challenges brought by the above problems and improve the counting accuracy of the algorithm.The specific research contents are as follows:1.Aiming at the problems of complex background interference and pedestrian scale change,a crowd counting algorithm integrating based on codec network structure and channel and spatial attention is proposed.The algorithm consists of two parts: encoding and decoding.In the encoding stage,the feature information of different depths and different pedestrian scales is extracted through the feature extraction network.In the decoding stage,repeated convolution,upsampling and other operations are used to recover the scene space information.At the same time,multi-level feature fusion is used to make up for the loss of detailed feature information caused by the deep network,and to overcome the problems caused by pedestrian multi-scale performance.the attention model is integrated into the decoding stage,and the weight is strengthened in the two dimensions of channel and space,to overcome background noise.The experimental results show that the algorithm proposed in this paper can effectively solve the problem of background interference and pedestrian multi-scale,and improve the accuracy and robustness of counting.2.To improve the utilization of spatial information in the scene and solve the problem of multi-scale and disorderly distribution of pedestrians,a dense crowd counting algorithm based on multi-level feature fusion is proposed.Adaptively solves the pedestrian multi-scale problem by using the multi-layer depth feature repeatedly.The algorithm as a whole is an inverted triangle structure,and the features of pedestrians at different depths are extracted through the backbone network,then the multi-layer structure is used to repeatedly fuse the features at different depths,to achieve the purpose of adaptive identification of multi-scale pedestrians.A loss function combining the global number of people supervision and pixel point supervision is designed to improve the utilization of spatial information in the scene and ultimately improve the model accuracy.The experimental results show that the algorithm has excellent accuracy and robustness in real scenes.Main innovations:(1)By using the encoding and decoding structure of the convolutional neural network and the attention structure of fusion space and channel,the adverse effects of pedestrian scale change and complex background interference on the counting performance of the model in dense scenes are reduced;(2)By reusing multi-layer depth scene features,multi-scale pedestrian problems can be solved adaptively;the combination of the global number of people supervision and pixel point supervision makes the utilization of scene space information and model accuracy improved.