| Crowd counting aims to estimate the number of people and the density distribution using the crowd characteristics in a given scene,and it plays an important role in many industries today as a video image processing technology.Especially with the growth of the world’s population and the development of urbanization,there are more and more application scenarios of crowd counting,which is of great significance to social security and crowd control management.Thanks to the continuous improvement of computer hardware and the rapid development of deep learning,considerable progress has been made in the study of crowd counting in recent years.This paper specifically analyzes the difficulties and challenges at this stage for different complex scenes,and mainly studies and solves problems such as background misjudgment and uneven density distribution in the scene.Three crowd counting algorithms based on convolutional neural network are designed,all of which significantly improve the counting accuracy and density estimation map quality,and verify its effectiveness and robustness on commonly used datasets.The main research and innovation are as follows:(1)In order to enhance the features of interest and reduce the background misprediction,a space-aware crowd counting algorithm based on the bidirectional supervision of the foreground and background is proposed.By fusing channel features and spatial features,highresponse features are emphasized from two dimensions,and the expressiveness of crowd features is enhanced,thereby improving the representation ability of the network.Finally,two decoding branches are used to supervise the foreground and background prediction respectively,which strengthens the network’s ability to distinguish between human head and background,significantly reduces the misjudgment of background noise,and improves the robustness of the network to complex backgrounds.(2)In order to enhance the expressiveness of human head features and then improve the performance of the model,a dual-task interactive crowd counting algorithm based on multilayer supervision is proposed,which improves network efficiency by supervising the distribution and counting results of crowd features at different stages.The early foreground enhancement module improves the discrimination of background noise by increasing the shallow network’s attention to the foreground.The counting layer emphasizes the characteristics of high-density areas and misjudgment-prone areas according to the contribution of different areas to the prediction results,effectively reducing missed judgments and confusing background misjudgments caused by insufficient foreground prediction.The different levels of supervision in the algorithm will work together to adaptively emphasize high-response features by using multi-layer semantic information,which greatly improves the model performance with minimal space cost.(3)In order to improve the performance of the network in sparsely distributed areas and further reduce background misjudgments,a foreground attention crowd counting algorithm based on background suppression is proposed.The algorithm significantly improves the prediction accuracy of sparse areas by emphasizing the response of foreground features to counting,and avoids the over-focusing of crowd response map weighting on dense areas.At the same time,the algorithm also significantly reduces the wrong prediction of the background region in the density estimation graph by penalizing the wrong prediction on the background in the density estimation graph,which is helpful for subsequent scene analysis work.A large number of experiments and analysis have proved the effectiveness of the algorithm in this paper.The algorithm in this paper not only improves the counting accuracy,but also significantly improves the quality of the density estimation map,which lays a good foundation for the subsequent scene analysis and crowd localization work. |
