Research On Target Tracking Algorithm With Rotation Estimation

In recent years,with the rapid development of information technology and the continuous enrichment of data acquisition methods,people have more and more demands for video and image data processing.Target tracking is an important technology of image and video data processing,and its related research has also become extremely popular.From the correlation filtering tracking algorithm based on traditional computer vision technology to the tracking model based on deep learning,the performance of object tracking algorithm has been greatly improved.However,,object tracking still faces the challenge of some complex situations,e.g.,background clutter,illumination variation,target occlusion and rotating and scale changes of moving objects,which seriously affect the performance of existing target tracking algorithms,and also complicate how to accurately track the target.This paper conducts in-depth research on the single target tracking algorithm of deep learning and correlation filtering,and proposes a target tracking model that can cope with the tracking challenge caused by target rotation,then applies it to the in-air handwritten character recognition system and achieves good performance.The specific research content includes:(1)In the object tracking task,full-convolutional siamese networks(Siam FC)tracker contains some tracking errors or inaccuracies caused by the rotation and deformation of moving objects.A Siam FC tracking algorithm with rotation and scale estimation was proposed in this paper.The proposed algorithm included location model and rotation-scale estimation model.First,in the location model,the tracking position was obtained by the Siam FC algorithm,and adjusted by rotation and scale information.Second,in the rotation-scale estimation model,the object search area was transformed from Cartesian coordinate system to logarithmic polar coordinate system,by which the rotation and scale transformations were became translation transformation.Then,the scale and rotation angle of the object were estimated by the correlation filtering technology.Finally,an object tracking model which can simultaneously estimate object position,rotation angle and scale change was obtained.The experimental results on the POT and OTB public data sets demonstrate that the performance of the proposed algorithm is superior to that of some common target tracking algorithms such as Siam FC,especially on video sequences containing rotation and scale changes.(2)In view of the fact that the real-time performance of the deep learning target tracking model is achieved by strong computing power and GPU acceleration and the kernel correlation filtering algorithm with better real-time performance requires very low computing power,we introduce the rotation and scale estimation module in the kernel correlation filtering model.The position and rotation-scale parameters of the target are estimated in the Cartesian coordinate system and the log-polar coordinate system respectively,by which we design a correlation filtering tracking model that can simultaneously estimate object position,rotation angle and scale change.Then,we apply it to a real need and develop a practical in-air handwritten character recognition system,which uses an ordinary camera as the image acquisition equipment,adopts computer vision and machine learning technologies,such as target detection,target tracking and text recognition,to record and reconstruct the written characters by detecting the fingertip and tracking its motion trajectory,and then recognize them by trained character recognition network.