Object Counting Based On Point-level Annotations

With the advancement of urbanization,a large population influx into central cities has caused a surge in population density,putting urban security to the test.Crowd gathering is prone to crowd trampling events.In this thesis,we study object counting algorithms to monitor the number of objects in target scenes using image data.The object detection methods can achieve object counting by counting the number of predicted bounding boxes.However,the object detection methods require box-level labeling during the training process.For dense scenes where targets are clustered,there are mutual occlusions between targets,and it is difficult to accurately annotate the boundaries and sizes of targets.This thesis focuses on exploring the object counting task based on point-level annotations.However,existing point-level annotation-based object counting methods predict the counting result of a single category in one inference process.In this thesis,we design a multi-class object counting method based on point-level annotations.We design a multi-class object counting task based on point-level annotations,organize multi-class object counting datasets,and design evaluation metrics.Multi-class object counting methods have broader application prospects than the single-class object counting methods.The existing point-level annotated object-counting datasets only provide single-class annotations,which cannot satisfy the training requirements of multi-class object counting methods.In this thesis,we construct multiclass object counting datasets based on point-level annotations.Next,we propose a dilated-scale-aware category-attention network.This thesis draws on attention mechanism and feature fusion techniques for algorithm optimization.Firstly,a category-attention module is designed to overcome the inter-category noise response.Due to the perspective effect,there are scale variations between targets at different distances.The scale variations profoundly affect the object counting performance.We further design a dilated-scale-aware module to fuse the detailed information and global information contained in feature maps of different resolutions to enhance the robustness of the counting methods to scale variations.We employ the proposed optimization modules to design the dilated-scaleaware category-attention network.Extensive experiments present that the proposed method equips with good object counting performance.This work effectively riches the application scenarios of existing object counting methods.