Research On Small Target Recognition Technology For Small UAVs

In recent years,small drones have been widely used in such fields as met eorological exploration,map mapping,border control,forest fire prevention and rescue,disaster monitoring,pesticide spraying,traffic control and geological s urvey.The UAV visible light image has a small target,a weak contrast and a complex background.The small target may be almost completely submerged in the noise.Under such a complicated background condition,it becomes very d ifficult to detect the small target.In order to solve the problems,the paper fir stly divides the small target image and divides the small target from the backg round.Then the SLIC fine segmentation is performed to segment the small tar get that is easy to hide in the background.Then the DBSCAN clustering algor ithm is used to eliminate some false targets.A small target detection method b ased on SVM is proposed.The main work of this paper includes:1.Segmentation of the small target.For the problem of small contrast of the small drone image,the light of the small target image is enhanced.Then t he small target image is filtered and the morphological transformation separates the small target from the background.The small object size is relatively smal 1,and it is easy to be hidden in the background.Then the small object image is subjected to SLIC fine segmentation.and the small object image is divided into super pixel units.The luma space of the super-pixel unit is used as the c lustering feature.The DBSCAN clustering algorithm is used to cluster the seg mented super-pixel units and eliminate some false targets.2.Feature extraction of small target.The small target size is relatively sm all,and it is very difficult to directly extract features such as HOG and LBP.The traditional method directly uses the mean,variance and information entrop y as the input of the SVM to train the SVM.Practice has proved that this m ethod can not deal with complex scenes.This paper extracts the characteristics of the mean,variance,and information entropy of the small target.After the f eature fusion of these small targets as the input of the system,an improved fe ature training method is proposed to train the SVM.3.Small target recognition algorithm migration.Transplantation verification was carried out for the applicability of the project,and the small target recog nition algorithm was transplanted to the DM642 platform.At the same time ta king into account the real-time requirements of the design,using the RF5 refer ence framework to create three new tasks,respectively,to achieve video captur e,processing,display tasks,the three tasks to cooperate with each other to ac hieve the detection of small targets.The experimental results show that the algorithm has a certain accuracy in complex scenes and meets the real-time requirements.