Research On Deep Learning Based Visual Object Tracking

Visual object tracking is an important branch of computer vision. Tracking task involves multi-domain knowledge, also accompanied with diversified interference. Deep network has much priori infomation, good generalization performance as well as rich hierarchical feature, which is suitable for making up manual features' problem of ralatively limited expression capability to improve tracking performance. Current deep learning tracking usually uses only top-level oupput to judge target state, ignoring net's middle features and other tracking methods. In that fully using net features and interacting tracking module more elaborately represents a promising tendency for raising tracking accuracy.To make fine-grained integration between deep learning and video tracking, two tracking algorithms are come up with, which are Deep Network Multiplexed Layer Feature Tracking, MLT, and Multiple Network Sub-block Tracking, MNT.MLT uses diversified network features, interactes the layer feature with tracking modules. It pyramidly uses the convolutional feature constructing appearance template, designs target specialized discrimination network to infer the target state, gets input layer saliency map inversely to differenciate interference factors and choose update strategy. All operations are completed during one propagation going forward and backword tightly and efficiently. It also proposes a three-stage template recovery strategy using features above. Experiments show that MLT can handle long time whole target occlusion and severe deformation to some extent, with high tracking accuracy and good robusteness.MNT merges deep learning and object tracking by splitting the target. Firstly sub-blocks are gotten by convolving image with Laplacian calculator. Then MNT uses middle depth network trained from large scale tiny images database to track sub-blocks, reducing computational complexity as well as raising precision. At last it updates network with long or short term samples. Experiments prove the effectness of MNT.