Research On Key Technologies Of Aerial Ground Object Detection In Complex Background

The urban population is increasing day by day,while the aging of population is getting worse and worse.In addition,the environmental pollution is intensified,while traffic congestion becomes more and more frequent.The disadvantages of the existing urban operation and management model are gradually emerging,especially under the severe situation of global epidemic.There is an increased demand for construction of smart cities.As the fundamental of smart city operation,the perception and understanding of urban information has attracted more and more attention from countries and regions.Traditional ground sensors are difficult to satisfy the needs of urban construction due to their poor flexibility and limited sensing range.Meanwhile,the development of UAV(Unmanned Aerial Vehicle)technology provides new solutions for urban information sensing.In recent years,aerial photography,as an important way of UAV Ground perception,has attracted the attention of more and more scholars.At present,great progress has been made in object detection in aerial images.However,the detection of dense and small objects in complex background is still challenging.Therefore,this paper conducted a series of innovative research on the detection of dense and small targets in aerial images under complex background,and proposed a series of methods for the difficulties and key problems existing in the task.The main work of this paper is as follows:1.In view of the problems that the existing detection models are difficult to adapt to the diversity of object scale,position and attitude,or the instability of the solution process due to the lack of reference information when solving the object position,this paper proposed a bag based sampling mechanism and fitness matching mechanism,which defines positive and negative samples according to the fitness between the objects and bags.In addition,a data enhancement algorithm based on weak supervised learning was proposed to solve the imbalance problem between foreground classes,which could generate more reasonable context information without additional manual labeling mask.2.Aiming at the problem of similar feature distribution of dense and small objects in aerial images under variable inclination angle,this paper introduced multi-instance prediction mechanism,and proposed MIP-BSSD,which uses similar features to detect multiple objects in one bag,This paper also proposed a new fitness function to solve the definition of positive and negative samples in multi instance prediction.In addition,aiming at the problem that the features of small objects are not obvious,this paper proposed a feature enhancement algorithm based on two-way fusion,which fuses the features with different receptive fields to improve the discrimination of features.For unreasonable false alarms,this paper proposed a false alarm removal algorithm based on outlier analysis,which constructs outlier confidence function without introducing additional information,and can effectively remove unreasonable false alarms.3.Aiming at the problems of solving the inverse of the atmospheric scattering model,and obtaining paired haze images and haze free images in practical application,this paper utilized the adversarial learning mechanism to make the deep neural network accomplish the approximate solution of the atmospheric scattering model in the process of game.4.Aiming at the problem of strong light interference in dynamic aerial scenes,this paper proposed a strong light interference suppression network named MSOCL-GAN.The image content is decoupled from the strong light interference by using the method of unpaired image translation.Meanwhile,the MSOCL loss was designed to make the network not only pays attention to the decoupling of the overall content of the image,but also pays attention to recover the detail information of objects.