Technology Of Uav Object Tracking Based On Computer Vision

With continuous development of technology,Artificial Intelligent(AI)has been widely applied to all kinds of fields.Computer vision,as the main direction of AI,also has grown rapidly in recent years which made its important branch--target tracking technology get more and more widely applications such as intelligent security and traffic,safety driving and video analysis etc.At the same time,because of the convenience of the aircraft deployment and broad range of the movement,combining with the target tracking method of UAV platform,it gives a huge advantage in detection and navigation.The target detection and tracking technology during navigational field has been widely applied to the military affair and civil area.It is quite successful so far and will continues to be one of the hot spot in the future.The target tracking technology of UAV can be implemented in different ways.For example,we can install a tracker in the target to realize the long-term tracking.But the tracking based on the vision is most intuitive and accessible to the human behavior.This paper is based on the computer vision for the target tracking research.The traditional ways of target tracking algorithm are mainly based on the manual features(such as LBP,SURF and HOG etc),for images features extraction,then use examples to train the classification which can distinguish the object and the background so as to complete the tracking process.Such type of method has a lot to improve in the UAV tracking area.Since the shooting of navigation videos are quite far away from some moving objects,as a result,some features like the resolution of the video imagine might be relatively low,the contrast of grayness of background and object is not obvious and borderline is obscure etc,which present a new challenge to the traditional target tracking algorithm.This paper is mainly to improve the traditional tracking algorithm to implement the real-time and long time tracking based on the UAV platform.Firstly,residual learning mechanism was used to enable deep learning methods to achieve end-to-end tracking,so as to improving accuracy and speed during tracking.Furthermore,In order to implement the real-time tracking,this paper mainly adopts the infrastructure of correlation filters to get target position quickly.Besides,this paper combine HOG and color feature for getting a more precise location.As for the fast motion problem,this paper also use the context learning method based on correlation filter to implement the tracking model.At the same time,since the size of the object appear ing in the UAV view varies a lot,this paper uses multi-scale variation strategy to get the most optimized selection for target area,thus the precision is guaranteed as well.Lastly,this paper combines with the keypoint matching method to relocate the object quickly so as to avoid the problem that the target area can not continue tracking due to occlusion reappearance.Finally,this paper not only use the traditional data-set OTB to do the data analysis,it also verifies more representative UA V tracking target data set UAV123.According to the experimental research,the method that this paper proposed is more advanced than the traditional method in terms of long-term and realtime tracking.