Research On Object Tracking Algorithm Of UAV Based On Deep Learning

With the continuous development of machine learning technology,deep learning algorithms are more and more valued by researchers.How to achieve high-precision real-time target tracking is an important topic in the scene of UAV target shooting.Due to the particularity of the aerial scene of the drone,the morphological changes,scale changes,and illumination changes of the target will result in unstable tracking performance and even loss of target tracking.The traditional UAV target tracking method is not good for the target tracking of scale transformation,and it is difficult to deal with the target of high-speed motion.It is not good for the target with excessive target displacement or occlusion between adjacent frames.Since the error of the traditional tracking system and the loss of the target are very serious,we propose a target tracking solution based on deep learning,which can play an important role in the field of UAV target tracking.In order to track and analyze ground targets,we have also developed an algorithm for recording target paths in real time.The algorithm can record the coordinate information of the tracking target in each frame of the image in real time.After comparing the performance of many deep learning algorithms,we finally adopt a SSD(Single Shot Multibox Detector)algorithm that directly predicts the bounding box coordinates and category confidence.The candidate targets are extracted.and then the Barthel coefficients of the target to be tracked and the candidate targets are calculated by combining the color histogram and the histogram of oriented gradients features,and the closest one is used as the matching target,and then the false tracking target is eliminated based on the Kalman filter.Experiments show that the proposed method is feasible,with high accuracy and good real-time performance.