Ensemble Object Tracking Based On Deep Learning

Visual tracking aims at training computer to simulate the ability of visual positioning as human.It could continuously track targets in moving scenes and is applied in the field of human-computer interaction,automatic driving,defense military,public monitoring and etc.Noways it has become one of hot research directions of computer vision.In the real world,the issues such as illumination variation,cluttered backgrounds,scale changing and occlu-sion in video pose great challenges for tracking tasks.In order to adapt to the variation of the target appearance,the mainstream tracking algorithms often use the pattern recognition method to establish a target appearance model with learning ability,and update the appear-ance model with the image frames acquired in the tracking process.Compared with the traditional pattern recognition method,deep learning has more powerful abilities of nonlinear fitting and self-learning,mining the potential principle inside the data.In recent years,the application of this technology has made breakthroughs in the research of visual tasks such as target classification,detection and segmentation.However,visual tracking tasks need to address pervasive tracking of non-specific targets,which are likely to be inconsistent with training data.Therefore,in the tracking application of deep learning,the on line update of the appearance model has become an urgent problem to be solved.Specif-ically,the research on this problem mainly has the following difficulties:Firstly,the data that can be provided by the tracking is limited.When the parameters of the neural network are updated,it is easy to fall into the local optimal.Especially,when the appearance of the target changes drastically,the neural network cannot adapt to the current state,resulting in ambiguity;Secondly,beside the first frame,the label of the training data depends on the accuracy of the predicted position from the previous frame,and the mistake label will make the deep neural network continuously update and bring the accumulated error,leading to the wrong estimation.To solve the above problems,we propose to employ the splitting mechanism of the tree structure for retaining the diversity appearance model,and the combination strategy of en-semble learning for the collaborative prediction of the target position.When the tracker detects the drastic variation of the target appearance,it searches the best matching node of the tree with the current target,updates the information of the node and splits it to store the updated model.The appearance model stored by the leaf nodes generated by the splitting mechanism has diversity,and m ore accurate prediction results can be generated by the rule of ensemble learning.The tree structure is designed to ensure the consistency of the training samples in each path under the condition of insufficient samples.The positive samples for model updating in the same path are continuous,so the ambiguity problem caused by the excessive difference of training data can be avoided.In order to improve the accurate prediction of target location further,in the update operation of nodes,we propose to adopt an update strategy based on active learning and manual data for the improved adaptability of the neural network fitting current data.This strategy emphasizes the diversity of training data.Active learning can actively select samples on the classifica-tion boundary during the training process,improve the learning ability and efficiency of the model,and avoid the model over-fitting to the easy-to-classify samples.In addition,the ar-tificial data is generated by the empirical distribution of the feature data,and the fine-tuning network with diversity is selected by maintaining the prediction accuracy,thereby making up for the predicted error of neural network caused by the missing training samples.Ac-cording to the proposed target tracking algorithm,this paper evaluates the performance of the algorithm based on the OTB and VOT platforms.Firstly,the contrast experiments show that compared with the mainstream algorithm,the proposed algorithm effectively improves the robustness of tracking,and solves the problems of illumination change,fast motion and occlusion in the tracking process to some extent.Secondly,it shows that the update strategy based on active learning and artificial data could improve the learning efficiency of deep neural network and enhance the accuracy of model prediction.