Research And Application Of Object Detection Based On Target Appearance And Geometric Modeling

Object detection is a challenging subject in the field of computer vision research,which is the basis of advanced visual tasks.Although it has been studied for decades,the performance of object detection is still inadequate facing the complicated real world.As a complex question of classification and location tasks,object detection is always stuck with the tradeoff between model capacity and computation cost.According to the geometrical variation,the target is divided into structured and unstructured object.The core of structured object detection is how to express the geometrical informa-tion of object.and model the geometric variation of structured object.For structured object detection,this thesis assumes that the geometrical variation of object is perspec-tive transformation.It expresses the geometry of object with point feature set and models the structured object detection algorithm with S-SVM classifier.This thesis proposes a pre-training and tracking algorithm to further improve the efficiency of structured object detection.Experimental results show that pre-training algorithm improves the discrim-inative ability of the classifier and tracking algorithm greatly improves the detection speed without losing accuracy.The core of unstructured object detection is how to express the region information and combine region proposal extraction and object classification into a single unified framework.For unstructured object detection,this thesis expresses the region infor-mation with data-driven feature and models the unstructured object detection algorithm with Faster R-CNN.This thesis proposes a region proposal fusion algorithm based on multi-level stimulus and further improve the detection efficiency of unstructured object-s.Experimental results show that the multi-level stimulus method enriches the feature abstraction ability and supplies the learner with better classification rules.Overall,this thesis analyzes structured and unstructured object detection methods and proposes the corresponding algorithm to improve the object detection efficiency under certain application scenarios.