Research On Ground Target Detection Algorithms On Airborne Platform

Target detection on airborne platform has always been an important research content in the field of military affairs.It is an important part of airborne photoelectric early warning,and airborne photoelectric investigation.Traditional target detection algorithms are widely used on airborne platforms,which are generally composed of three parts: region selection,feature extraction,classification and regression.For example,the classical strategy based on target template matching is very popular,which has many advantages including easy algorithm implementation,fast detection speed and excellent detection results.However,it has some obvious shortcomings,such as the need for a certain prior knowledge is indispensable,and once a frame fails to be detected,its update template is vulnerable to background pollution.Therefore,a more complex template update strategy in these algorithms is required.However,the imaging environment of aerial image is complex and unpredictable in many cases.The imaging quality of the target will be affected by the influence of weather,the changes of illumination and the shaking of fuselage.At the same time,it is easy to cause some deformation of the imaging target due to the rapid movement of the airborne platform itself or the fast movement of the target itself.Traditional algorithms still do not achieve better robustness to the target.Therefore,this paper studies the detecting algorithm of the small target in aerial photography based on deep learning.Depth features show great advantages in changing factors such as object deformation,illumination variation,and background interference,compared with traditional feature extraction methods such as harr,sift and hog.But there are inevitably some shortcomings,for instances,the computational complexity is increased significantly and a lot of computational is needed urgently.Modern aerial camera target detection requires not only excellent detection accuracy but also high detection speed.The main work of this paper focuses on the following aspects.(1)In order to cope with complex environmental changes,this paper proposes a series of data preprocessing measures,including common data enhancement strategies such as rotation,cropping,brightness,color,etc.to enhance the robustness of the model.It is often faced with interference from clouds and bad weather for target detection on airborne platforms,Therefore,the defogging algorithm is specially studied to reduce the loss of imaging information in the physical environment and illumination,and to enhance the stability of the algorithm.(2)This paper proposes a method for multi-layer feature fusion for target detection in order to improve the detection accuracy of aerial targets on airborne platforms.The shallow feature extracts the edge and contour information of the target,which is beneficial to achieve accurate target positioning.At the same time,the deep feature extraction is the semantic and environmental information of the target,which is conducive to the accurate recognition of the target.Therefore,making full use of the deep and shallow features of the target can better capture more information of the target,which make up for the deficiencies of the YOLO algorithm for small target detection.And the improved non-maximum strategy is used to improve the impact of target overlap in the final output target.The detection accuracy reaches 0.63 mAP on the self-made data set.(3)In order to reduce the model storage requirements and improve the detection speed,this paper applies the singular value decomposition strategy to the feature map,in which the shallow feature map compression effect reaches 80% or more,and the deep feature map compression effect reaches 30% or more.At the same time,the whole model detection speed can achieve 18 fps.