Research On Object Tracking Algorithm Based On Deep Learning

The main purpose of video object tracking is to track any target that will change over time,and the ground truth of the target will only be specified in the first frame.The difficulty lies in how to effectively use very limited training data to build a tracker that can adapt to difficult scenes such as similar background,fast motion and occlusion.In recent years,the Siamese tracker of the end-to-end network structure has achieved excellent performance in object tracking.Because of its excellent performance and efficiency,it has been favored by a large number of researchers due to its excellent performance and simple network structure.This tracker does not perform well for difficult scenes such as fast motion,occlusion,and the accuracy of the Siamese tracker when tracking is extremely dependent on the function of the cosine window,which reduces background interference.The tracker usually selects the search area based on the target position in the previous frame.When the target moves quickly,it may cause the cosine window penalty mechanism to become unbalanced,thereby exacerbating the boundary effect.Secondly,the tracker usually constructs an online learning model under the assumption that the target response has a Gaussian distribution centered on the target.However,this assumption is not feasible when encountering other targets or background noise constantly interfering with the target.In this case,we can observe that the higher sub-peaks are generated on the tracking response map,indicating model drifts.Finally,Model object trackers largely rely on the online learning of a discriminative classifier from potentially diverse sample frames.However,noisy or insufficient amounts of samples can deteriorate the classifiers' performance and cause tracking drift alterations such as occlusion and blur-ring can cause the target to be lost.To address these issues,this paper mainly focuses on building a robust object tracker.Focus on the cosine window imbalance penalty,fast motion,sub-peak suppression,robust model learning,occlusion blur and other difficult problems in object tracking,and proposed solutions.Our main contributions can be summarized as follows:1.In order to solve the boundary effect of the tracker caused by fast motion,we propose a positioning network to use the abundant motion information between adjacent frames to accurately track the target.We apply the optical flow estimation network to predict the movement trend of two targets in adjacent frames.This is a good solution to the difficult scene of fast-moving tracking.Making full use of the motion information also makes our tracker more robust.It further eliminates the influence of the boundary effect caused by the uneven penalty of the cosine window,and the performance of the tracker can be greatly improved.2.We introduce the Kalman filter method into the Siamese tracker to obtain the target's trajectory information,which is conducive to our tracker's robust tracking of difficult scenes such as occlusion and fast motion.We further propose to adaptively combine the response scores under different features with a certain weight.Making full use of the features of multiple layers also allows our tracker's accuracy to rise further.3.Based on the model drift of existing trackers,we propose a sub-peak suppression tracker(SPSTracker)to reduce the model drift by suppressing potentially large subpeak values online and highlighting the maximum peak response value.To achieve stable tracking of the target tracker.We propose a simple and efficient peak response pool(PRP)module to aggregate and align the discriminative features,and design a boundary response truncation(BRT)module to reduce the variance of feature responses.They are applied to any target tracker in a plug-and-play manner.4.Finally,we improve the robustness by finding out the most beneficial input training samples for the tracker,and removing noisy samples as much as possible.Specifically,a simple-yet-efficient Significant Sample Selection(S3)strategy is proposed.The method relies on(aleatoric)uncertainty adjustment,with the uncertainty estimation embedded into the regression branch of the model through a carefully designed loss function.To further improve the robustness of the tracker,a “Mixedfeature” data augmentation method is proposed and applied to the classifier.The method consists in adding samples perturbed through simulated occlusions and blurring to the training data set,and achieves significantly in-creased performance in the event of actual target occlusion or blur.Furthermore,we apply deformable convolution to the backbone of the network and perform an end-to-end training.