A Visual Object Tracking Method Via CNN And Optical Flow With Online Learning

Object tracking is a semi-supervised learning problem with insufficient training samples.From the point of implementation,it is defined as the embodiment of estimation for the motion trajectory in a scene.In the same words,it is considered as detecting the target so as to distinguish the tracked target from the back ground in every given frame of the video stream.In the field of military navigation and security of public life,it is widely used to take the place of human beings.In this paper,we propose a method based on Bayesian method,convolutional neural networks(CNNs)and Pyramid Lucas-Kanade optical flow(PLK optical flow),which is called COB method,for object tracking problems.With the idea of track-by-detect,the COB is composed of a detector based on CNNs,a tracker based on PLK optical flow and integrator based on Bayesian method.The CNNs based detector is cascaded after the PLK optical flow based tracker.After that,the results from the tracker and detector are fed to the Bayesian method to integrate the final result of the COB method.The contributions of the paper can be summarized follows.Firstly,an adaptive integrating method is used to solve the over-fitting problem caused by insufficient training samples,so as to reduce the influence of over-fitting.The adaptive integrator introduces a challenging mechanism between tracker and detector in order to compensate errors from tracker and detector.On one hand,we use the result to clear the cumulative error of the tracker;on the other hand,the result from the tracker compensates drifting from detector.Secondly,considered the practical applications,an optional online learning mechanism is also introduced by the integrator.One can enable the online learning mechanism to get better performance and sacrifices the real-time performance.Or the online learning mechanism is disabled to enhance the real-time performance.Thirdly,a new network structure is also come up with so as that the network is able to make real-time classification with a good performance.Besides when the network is trained,some regularization method in deep learning is introduced to the network,in order that the network is robust to the blind spot of classification,and the probability of misclassification is reduced.In this paper,an object tracking method called COB is proposed.The method integrates the CNNs and PLK optical flow into the frameworks of Bayesian method.In COB,PLK Optical Flow based tracker is used to make prediction of the position according to the position of next frame,which is able to mine the information between two adjacent images while the CNN based detector is used to detect the position of the target,whi ch is used to mine the information in current frame.The COB is implemented including the COB with online learning and COB without online learning based on the CVPR2013 object tracking data set.We have conducted extensive experiments to investigate the feasibility of our COB method,the statistical results show that our COB method always obtains comparable perform compared with the previous methods,especially COB with online learning.