Vision-based UAV Target Recognition And Tracking

With the rapid development of drone technology,target recognition and tracking based on drones has quietly become the focus of research in the field of computer vision.In fact,there is a need to use drones in the air to inspect the area in the military and in life.However,in the existing algorithms,it is weak to deal with the high-altitude recognition and tracking targets,because the target may have defects such as small size and missing texture information.Therefore,this paper proposes a structure based on vision-based UAV target recognition and tracking,with the aim of improving the detection and tracking performance of small pixels at high altitude.The subject comes from the actual needs of the military regional inspection.Firstly,we study various methods of target detection and tracking,analyze its performance and inferiority,and find that the regression-based detection algorithm can take into account the requirements of speed and precision,and perceive that the performance of the tracking method of related filtering is even better.The YOLOv3 and KCF algorithms are selected according to the application background.Secondly,it solves the problem of image distortion caused by abnormal weather and sudden changes in light.The image quality is enhanced by color image graying,mean filtering,histogram equalization and image correction,which reduces the difficulty of subsequent recognition and tracking.Then the YOLOv3 detection algorithm based on cluster analysis anchor box and prediction scale is studied.First,a brief description of the basic neural network and the YOLOv3 detection algorithm is given.Then,according to the application scenario of detecting the target at high altitude,the anchor box is clustered and analyzed according to the characteristics of the collected data set to obtain the corresponding size,so as to improve YOLOv3.Then improve the network for the characteristics of the target,cancel the prediction of 3 scales,and only use the feature map of 4 times downsampling to identify the target.And to prevent the gradient from disappearing,the final DBL unit is improved as a residual unit.Then a KCF tracking algorithm based on motion prediction,scale adaptation and detection adjustment is proposed.The KCF algorithm adopts the methods of cyclic matrix,kernel function and Fourier transform domain to improve the accuracy of tracking in the case of small calculation.However,due to the fact that the target motion speed is too fast,the direction changes frequently,and the target size changes greatly in the application scenario,the tracking failure will be caused.Therefore,this paper presents a KCF method that adds motion prediction,scale adaptation and detection adjustment.The designed tracking algorithm can predict the position of the next frame target by the trend of the previous frame motion.And a variety of scales are used,so that the target can still be tracked in case of a sudden change in target size.Moreover,the algorithm can also adjust the tracking result according to the detection result.Finally,simulation experiments are carried out according to the proposed algorithm structure.First,the data set is extended according to the application scenario of the high-altitude detection target,and then the algorithm training data set is written,and then the training index is evaluated according to the weight file.The results show that the performance of the high-altitude detection target is superior,the mAP reaches 89.06%,and the tracking position error is less than 7 pixels on average,and the tracking effect is good.