Object Detecton And Recognition Methods On Aerial Images From Low-altitude UAV Based On Deep Learning

The UAV,a kind of new combat platform,has played a huge role in several local wars and non-war military operations.The use of UAV combined with artificial intelligence algorithms for reconnaissance and situational awareness has been widely concerned and further studied by military powers.In recent years,the object detection based on deep learning has been continuously developed,and has gradually overcome the detection accuracy bottleneck of the traditional object detection algorithm.The object detection based on deep learning has become the mainstream research direction in the object detection field.Object detection on aerial images from UAV is a sub-topic of object detection,where deep learning also becomes popular gradually.Since the aerial image size is large,the target scale is variable,and the shooting angle is different from the conventional shooting angle,the object detection method used on general images is not suitable for aerial images.Therefore,it is necessary to propose a detection method suitable for aerial image based on the general object detection method.Firstly,this paper proposes an aerial image dataset construction scheme to solve the problem that the number of aerial image datasets is small and the applicable scene is single.According to the aerial photography system of public UAV,the attributes of images and objects in the aerial image dataset are analyzed.Since the method of acquiring aerial images is limited,four methods are proposed.Then the image is processed by the tool called Dark Label.Finally,a reasonable method is proposed to organize large volumes of data.Secondly,the gradient clustering SSD(Single Shot MultiBox Detector)model is proposed for the problem that the detection speed of methods based on deep learning for aerial images is slow.The model uses the gradient information to filter out the redundant regions in the aerial image,which increases the amount of computation of the model.Thus,the model improves the detection speed.The final test shows that the detection speed of the proposed model is increased by more than 60%,but the detection accuracy is reduced,compared with the classical model.Finally,a multi-model fusion structure is proposed for the problem that the gradient clustering SSD model is not accurate.The parameters affecting the detection accuracy in the gradient clustering SSD model are analyzed firstly,including the significant grid point threshold,Gaussian parameters and detection model.The optimal parameters are selected by calculation.The advantages of several models are determined by the analysis of the model detection results.The multi-model fusion structure combines with the advantages of parameters,and achieves good results overall.The final test shows that the multi-model fusion structure maintains the advantage of detection speed,and the detection accuracy is improved by more than 20%,compared with the classical algorithm.