Research On Detection Method Of Small Objects In UAV Image Of Cluttered Scene

UAV(Unmanned Aerial Vehicle)surveillance has the advantages of high productivity,low cost,broad searching area and flexible mobility,overcoming the shortcomings of traditional surveillance with stationary camera.The combination of vision technology and UAV can not only improve the intelligence of a UAV monitoring system,but also extend its various applications.However,the quality and cluttered background of images taken in the distance make it difficult to detect the interesting objects using the classic algorithms.The issues of the size variation,aspect ratio and class imbalance of objects are studied here,as well as the detection of small,dense objects in images,taken by UAV camera.An algorithm framework that combines shape perception and small sample amplification technology is proposed to detect the objects of various size,unbalanced categories in UAV images.The augmentation of small objects based on random sampling and filling is designed to accommodate to the issues of the various numbers of different objects in UAV images and the imbalance of their categories in the image dataset.Meanwhile,a feature pyramid module based on shape perception is designed to solve the problems of the size change and the various aspect ratio of objects in the UAV image,which cause the difficulty to characterize the multi-scale shape features with a deep network model.It obtains more powerful semantic and location information through cross-level feature fusion,and performs two-dimensional down-sampling operations on the new feature map to model objects with different sizes and aspect ratios,overcoming the limitation to predict objects from a single feature map.The above modules are tested on an anchor-free object detection network,and the ablation experiments were carried out with different backbone networks.The experimental results shows that the modules improve the accuracy and speed of object detection.In order to solve the problem of small object detection caused by the diversified shooting angle of drones and the long distance,A density map guided object detection algorithm is proposed to detect the small objects in UAV image raised by the shooting angle and distance.Considering the objects size and their distribution characteristics,the prior knowledge of the clustering effect of small objects in the image domain is used,and a density map generation method based on the object category is introduced to design a density map estimation module and predict the object with the density map module.The density information is used to crop the high-density area where the small objects gathered,and adaptively zoom in the cropped area.The object detection result of the cropped high-density area image is merged into that of original input image.Experimental results show that the proposed method can significantly improve the detection accuracy of small and weak objects.This thesis studies the small object detection of UAV images of cluttered environments,and proposes an algorithm framework,which combines shape perception and augmentation technology of small sample,and a density map guided algorithm framework,respectively.These frameworks may be applied to the research areas of urban planning,traffic monitoring,and disaster relief,and so on.