Research On Key Techniques Of Personnel Search Based On Aerial Images

In recent years,UAV has become a new member of aerial photography equipment carrier,and it is closely combined with communication and artificial intelligence,providing many traditional problems and new solutions,among which personnel search is one of them.Searching for people,which often occurs throughout the ages,can be divided into manual direct search and tool assisted search,which belongs to the latter.This search method is mainly used in the search or patrol of personnel in a specific area,for example,the search for lost "donkeys" and the patrol to check whether there are people staying in the dangerous area before blasting operation.Compared with direct manual search,this method has the characteristics of high efficiency,security and versatility.At the same time,with the popularization of electronic technology and 5g communication,it will be more widely used.In this paper,training set expansion and enhancement technology,super-resolution and target detection are combined,and contour judgment optimization strategy is used to achieve the balance of detection rate and efficiency:1.The shooting angle of aerial image is from top to bottom,and the distance from the lens to the target is far.These two factors bring the particularity of aerial image.As for the identification of people in aerial images,the most important feature is that "dual motion"(both UAV and human can be moved)produces the changeable image of people.According to the characteristics of UAV aerial image,when establishing training set,the full coverage of sample type needs to be realized by various sample expansion methods.In this paper,in addition to the traditional methods of sample expansion of exposure rate and color saturation,we also use rotation expansion,image atomization fuzzy expansion and generation of style transfer image expansion training set through anti network.2.In the process of aerial photographing,because of the difference of the height of the aerial camera,the angle of photographing and the posture of the subject,the image size of the target in the aerial photographing image is quite different.In order to improve the recognition rate,cross layer fusion is added to the feature fusion part of the target detection algorithm.Cross layer fusion can reduce the disappearance of key features in the process of feature extraction and scaling between different sizes of feature images.At the same time,the anchor frame prediction in the target detection algorithm is changed to the center point prediction.Anchor frame prediction is to preset different size recognition frame in the characteristic image of the image to be detected,but because the size of the target in the aerial image changes greatly,the size distribution of the preset frame is wide and difficult to predict.The center point prediction is the center point of the target imaging,and the prediction area is determined according to the change of the characteristic distribution near this point.This method does not need to estimate the image size of the target,and is suitable for aerial image recognition.3.In aerial images,human imaging is usually relatively small,and the detection rate of direct application of target detection algorithm is relatively low.Firstly,the super-resolution of the image is used to image the target,then the number of pixels is expanded,and then the target detection algorithm is used to identify the target.However,the super-resolution computation is large and takes a long time.Although it is not accurate in RGB image,it can be excluded from the classification.The contour of human imaging in aerial image has the characteristics of perimeter and area,which can exclude that most areas in the image are uninhabited,so it can directly carry out superresolution for areas that may be manned,and then apply target detection and recognition to these areas to achieve the collaborative optimization of efficiency and detection rate.4.The valuable information in aerial images is easy to be interfered by weather factors,especially the haze weather,which will greatly reduce the detection rate of target detection algorithm.In order to reduce the detection rate of targets in light haze weather,this paper uses the training set to include the aerial image data of people in haze and the image defogging algorithm to reduce the impact of haze weather on the detection rate of targets.Two forms of defog algorithm are given.When the relevant parameters of the haze degree in the search area are known,the scattering equation can be directly used to defog the image;when the relevant parameters of the haze degree in the search area are not known,the corresponding relationship between the target object trained by the anti network under the condition of no fog and fog can be generated to defog.Infrared thermal imaging is different from visible imaging because of the thermal radiation released by the imaging information source target,so this imaging method is not affected by the light and haze weather factors.The combination of visible light and thermal infrared imaging can improve the stability and robustness of the rate.5.In a specific scene,people may be in a specific object(such as a car),this kind of situation is difficult to be solved by the method of directly identifying the human image through object detection.In this kind of scene,we need to determine the object related to human first,and then identify it according to whether the state change of the object caused by human existence appears.Because the recognition of state change is the recognition of change process,most of the cases need to be based on the data obtained by target detection for analysis and judgment.In this paper,the types and methods of state change of visible light and thermal imaging are discussed.The combination of SVM algorithm and scoring method can reduce the judgment error caused by the fluctuation of target detection data.