Algorithm Study On Object Tracking Via Object Proposal And Deep Learning

Visual tracking is a main research in couputer vision,and it has extensive applications on intelligent video surveillance,human computer iteration,robotics,etc.It has attracteda large number of international and domestic academics.But,how to design a robust visual tracking algorithm still need much effort due to the highly complicated factors.Deep learning have make much progress in recent years,consequently,deep learning-based trackers have make good performance.But,dut to lack of training samples and offline pre-training is time-consuming,the learned generic features may less beneficial to online tracking.Moreover,the real-world data is with significant noiseand learned feature may have some task-irrelevant pattern,which will affect the performance of visual tracking algorithm.According to above-mentioned problem,we proposed two novel of visual tracking algorithm.The main research contents are as follows:1)We propose a robust visual tracking algorithm based on object proposal and PCA convolutional neural network.Deep learning-basedtrackers have been incrementally popular in the past few years.But the biggest bottleneck is lack of training samples.The offline pre-training is time-consuming,and the learned generic representation may be less discriminative for tracking specific objects.In order to solve above-mentioned problem,we propose an anline discriminative tracking method based on PCA convolutional neural network without large-scale pre-training,the filter banks in convolution layer are learned by PCA algorithm.A soft-thresholding operation is employed to the learned features to generate a sparse features representation.In order to further improve algorithm performance,we incorporate an object proposal mechanism that uses informative edges to score candidate samples,and design a multiple cues fusion method to choose the optimal target.Extensive results on the wdely used online tracking benchmark validate the robustness and effectiveness of the proposed tracker.2)We propose a robust visual tracking algorithm based on object proposal and point-wise gated convolutional deep network.The performance of visual tracking algorithm mainly depends on the feature representation(including feature learning and feature selection)of a target.In object tacking,a common assumption that video is without noise,which is too strict,especially with real-world data.Most tracking methods may fail if no good raw features.In order to solve this problem,we propose a novel visual tracking method via a point-wise gated convolutional deep network.Through a gating mechanism,the deep model can learn and select features together.Moreover,to alleviate the drift problem,we combine an edge box-based object proposal to improve the accuracy of the algorithm.Extensive results on the widely used online tracking benchmark validate the robustness and effectiveness of the proposed tracker.