Multi-Task Learning With Scale-Aware For Crowd Counting

With the rapid increase of the world's population and the acceleration of urbanization,the phenomenon of people gathering is very obvious.The high density of crowds not only leads to potential security risks,but also hinders the normal development of the city.As a direction in the field of intelligent monitoring,crowd counting has become a research hotspot in the field of computer vision.Effective crowd counting algorithms are of great significance for video security monitoring,traffic congestion control,urban planning and other fields.With the huge development of deep learning,convolutional neural networks have greatly improved the performance of different fields in recent years.Thus,it motivates researchers to employ convolutional neural networks to improve the performance of crowd counting.Due to the influence of crowd background noise and scale changes,the model is difficult to adapt to the complex crowd environment and achieve good counting results.In order to effectively solve the problems encountered in crowd counting,this thesis proposes a novel model named multi-task scale-aware crowd counting.By designing the scale aggregation module,the features of different scales are extracted,and these scale features are aggregated,so that the network can adapt to the scale variation of the crowd,greatly improving the model's feature representation ability and scale diversity,and effectively alleviating the scale variation problem.In view of the problem that background noise interference affects the counting accuracy in crowd counting,this thesis proposes to utilize segmentation-based methods to segment crowd features and background noise to remove the characteristic interference of background noise,effectively reducing the false count of background noise.Based on multi-task learning,this thesis introduces a crowd density-aware module to classify different crowd density levels.In the process of counting,through the classification of the crowd density level and the fusion of the classification information,the error of the crowd count is more effectively reduced.This thesis uses four common publicly available datasets,i.e.Shanghaitech,UCFQNRF,UCF?CC?50 and Mall,for experimental evaluation.Abundant experimental results show that the performance of the proposed MTSA model is better than the stateof-the-art methods.