UAV Perspective Vehicle Retrieval In Complex Scenes

Vehicle retrieval technology is a technology that uses computer vision algorithm to judge whether there is a specific vehicle target in the image to be measured.The technology has important application value in traffic planning,military reconnaissance and environmental detection.In addition,due to its small size and strong mobility,UAV is suitable for video and image collection in areas where information collection is difficult,so it has been applied in various civil and military fields.It is of great research significance to combine UAV and vehicle retrieval technology to study the UAV vehicle retrieval problem under different geographical and weather conditions captured by UAV airborne cameras.The research content of UAV vehicle target retrieval in complex scene includes three tasks:calibration of target candidate box in complex scene,accurate and fast retrieval,and construction of efficient target retrieval database.In the actual application environment,affected by UAV flying altitude,shooting Angle and background noise,the UAV vehicle target retrieval effect in complex scenes is poor and the missed detection rate is high,and the positioning is not accurate enough to meet the actual demand.Around these requirements,this thesis puts forward a kind of unmanned aircraft vehicle target detection method based on dynamic candidate box,solve the UAV aerial image target detection of the target class imbalance caused by the default bounding box and the target real box surrounded the deflection Surrounded by box high degree of overlap between lead to leak and inaccurate positioning problem.At the same time,aiming at the problems of complex background and large scale difference of target and difficulty in feature extraction in UAV target detection,a feature extraction network for vehicle target detection is designed.Finally,in order to ensure the construction of efficient target retrieval database,a closed-loop UAV perspective vehicle retrieval network is proposed.The specific research results are as follows.A dynamic candidate box extraction method is proposed.A dynamic candidate box is designed for UAV vehicle retrieval task in complex scenes.By introducing retrieval into target detection process,the candidate box is more suitable for the actual size of target,which improves the accuracy of location and reduces the rate of missed detection.A feature extraction network for vehicle target detection is designed.The feature extraction backbone network designed in this thesis draws on the design ideas of deep residual contraction network and Inception V3,integrates the attention mechanism,residual contraction network and inception structure,and constructs the feature enhancement module(FEM).It is demonstrated that this feature extraction network improves the accuracy of image classification to 98.9% compared with other feature extraction networks with the same number of parameters as Res Net-34.A vehicle retrieval network based on closed loop is constructed to solve the problem of constructing efficient target retrieval database,which can further improve the positioning accuracy of dynamic candidate box and save storage space and further reduce the retrieval time in the case of massive databases.