Research On Planar Object Tracking Algorithm Based On Optical Flow And Fully Convolutional Neural Network

With the continuous improvement of computer equipment and the continuous development of computer vision technology,as one of the core technologies of computer vision,planar object tracking technology has become a hot field of academic research in computer vision,and has a wide range of aspects in augmented reality,surveillance,robot vision,etc.Although the planar object tracking technology has made great progress in recent years,under the influence of external conditions such as size scaling,rotation,perspective distortion,etc.,the tracking recognition of planar objects may appear to varying degrees of loss or delay.To this end,this paper conducts an in-depth study on the tracking algorithm of planar objects:(1)Considering that in practical applications,the tracking algorithm needs to face complex and ever-changing environments at all times,and most of the current planar object tracking data sets are taken in indoor environments,this kind of image sequence lacking a complex and variable background cannot simulate the field well complex and changeable environment.Therefore,this paper proposes a data set for planar object tracking,which contains 40 objects,a total of 280 image sequences collected in the field with annotations.Each object contains 7 motion states: scale transformation,rotation,perspective distortion,motion blur,occlusion,moving out of view,and unconstrained motion.In addition,in order to reduce the workload of manual labeling,a semi-supervised labeling framework is proposed to assist manual labeling,16 advanced planar object tracking algorithms are evaluated using alignment errors,and the performance of each algorithm is analyzed.(2)The planar object tracking algorithm based on key points can be well adapted to complex scenes.The SIFT-based planar object tracking algorithm,as a classic method in this kind of method,has a good tracking effect.However,there are still many objects with wrong matching points in this method,which affect the tracking accuracy.On this basis,considering that the input image sequence has timing,the optical flow method can make good use of this timing information.Therefore,the optical flow method is integrated with the SIFT-based planar object trackingalgorithm,and a planar object tracking algorithm based on key points and fused optical flow is proposed.Experiments show that the planar object tracking algorithm based on key points and fused optical flow has better performance than the existing planar object tracking algorithm.Under the constructed plane object tracking data set of natural scenes,when the correction error threshold is 5,the tracking accuracy of the proposed method is improved by 3.8% compared with the classical planar object tracking algorithm.(3)Although the deep learning method of local feature matching applied to the planar object tracking task can achieve good results,the speed is limited by the number of key points in the input image.Therefore,this paper proposes a planar object tracking algorithm based on fully convolutional neural network,which can mitigate the impact of the number of key points on the tracking speed.The output feature map of the network has the same scale as the input image.During the training process,due to the extreme unevenness of the positive and negative samples in the training process,the network training is difficult.In order to alleviate this problem,the most difficult negative samples are sampled using the difficult sample sampling method,and the sampled positive and negative samples are sent to the loss function to train the network.The experimental results prove that compared with the planar object tracking algorithm based on deep learning,the proposed algorithm improves the proposed method by 4.1% compared to MATCHNET when the correction error threshold is 5,and is not affected by the number of key points in the input image.The tracking speed is limited and has a great advantage in speed,and the method also achieves good results in the general planar object tracking task.