| With the development of economy and urbanization,the use of video surveillance in daily life is becoming more and more popular.Traditional video data analysis methods are difficult to handle the huge amount of video data generated daily.In recent years,with the development of deep learning and artificial intelligence technologies,scholars have gradually applied deep learning techniques to the field of vision and have made significant breakthroughs in the fields of face recognition,target detection,image classification and crowd counting.Among them,crowd counting techniques can describe and estimate crowd density distribution information in video,which is especially important in today’s increasing population size.However,due to the problems of occlusion,variable scale,complex and variable scenes and difficult access to labeled data in realistic dense crowd scenes,the existing crowd counting methods are still not good enough to solve the above-mentioned problems.This thesis conducts research on the above-mentioned problems and proposes solutions from different perspectives,and the main research work and innovation points should or are as follows:To address the problems of scale variation and occlusion in complex crowd scenes,this thesis proposes an adaptive scale-guided network-based crowd counting method,which fully exploits the multiscale information in images through a multi-depth dilated convolutional adaptive fusion module,and uses an attention-guided fusion method to fuse context features layer by layer to effectively combat the scale variation and occlusion problems.This thesis also incorporates a depth-supervised approach to design a fine-tuning module to further improve the crowd counting performance of the network.For the video crowd counting scenario,this paper proposes a multimodal-based video crowd counting method.Unlike other methods that extract temporal features between video frames implicitly through LSTM and so on,this thesis uses optical flow to extract motion and temporal information in video explicitly,and designs modal sharing branches and information interaction modules to obtain the complementarity and correlation of features between modalities,so as to improve the robustness and accuracy of the algorithm.To address the problem of difficulty in acquiring labeled data in dense crowd scenes,this thesis proposes a weakly supervised and meta-learning based crowd counting method with few samples.This thesis is the first attempt to combine meta learn and weakly supervised methods in the field of crowd counting,and aids the adaptation process of meta learn methods in new scenes through weakly supervised tasks,so that the model can complete the scene adaptation process using a small number of weakly labeled images when facing new scenes.In this thesis,the above algorithms are evaluated in several publicly available datasets and ablation experiments are designed for each component of the algorithms.The experimental results show that several methods proposed in this thesis have good performance. |
