| As the demand for public security continues to increase,research on intelligent crowd density estimation algorithms has received greater attention.However,in practical application deployment,intelligent crowd density estimation algorithms face difficulties in cross-scene knowledge migration,scene knowledge self-learning and scene density prediction stability.Therefore,this dissertation addresses the problems above and improves the existing intelligent algorithms.The main contents include:1)Cross-domain Crowd Counting based on Self-supervised Learning.A two-stage domain adaptive method based on self-supervised learning is proposed to solve the problem that the performance of supervised learning methods is degraded when the data samples are insufficient.In the first stage,the method uses multi-scale feature response branches to more robustly learn cross-domain knowledge and reduce prediction inconsistencies across different scenarios.In the second stage,the trained model is used to generate robust pseudo-labels for the target domain,and the entire model is retrained with the pseudo-labels to enhance the adaptability of the target domain.Various experiments on the synthetic dataset GCC and three real public datasets validate the usability of our proposed method with higher accuracy.2)Explicit Invariant Feature Induced Cross-Domain Crowd Counting.Crossdomain crowd counting has shown progressively improved performance.However,most methods fail to explicitly consider the transferability of different features between source and target domains.In this dissertation,we propose an innovative explicit Invariant Feature induced Cross-domain Knowledge Transformation framework to address the inconsistent domain-invariant features of different domains.The main idea is to explicitly extract domain-invariant features from both source and target domains,which builds a bridge to transfer more rich knowledge between two domains.The framework consists of three parts,global feature decoupling(GFD),relation exploration and alignment(REA),and graph-guided knowledge enhancement(GKE).In the GFD module,domaininvariant features are efficiently decoupled from domain-specific ones in two domains,which allows the model to distinguish crowds features from backgrounds in the complex scenes.In the REA module both inter-domain relation graph(Inter-RG)and intradomain relation graph(Intra-RG)are built.Specifically,Inter-RG aggregates multi-scale domain-invariant features between two domains and further aligns local-level invariant features.Intra-RG preserves task-related specific information to assist the domain alignment.Furthermore,GKE strategy models the confidence of pseudo-labels to further enhance the adaptability of the target domain.Various experiments show our method achieves state-of-the-art performance on the standard benchmarks.3)Global Representation Guided Adaptive Fusion Network for Stable Video Crowd Counting.For the video crowd data,a global representation guided adaptive fusion network is proposed to explore the global spatio-temporal consistency of video sequences.The method establishes a robust long-term spatio-temporal representation between consecutive frames to guide the density estimation of local frames,thus alleviating the prediction instability problem caused by background noise and occlusion in crowd scenes.Experiments have shown that a 10% performance improvement is obtained compared to existing methods. |
