Visual Tracking Based On Deep Convolutional Neural Network

Online single object tracking is an important subject in artificial intelligence and computer vision,which is widely applied to video detection and surveillance,medical image understanding and analysis,human-computer interaction to name a few.In a typical scenario,an unknown target specified by a bounding box in the first frame is to be tracked in the subsequent frames.It is always a challenging problem to develop a robust tracking algorithm,with some factors such as severe occlusion,background clutter,pose and scale variation.In this thesis,we put forward a robust algorithm with deep convolutional neural network.The very deep neural networks have achieved favorable performance in image classification,object detection,etc.However,simply stacking more layers could suffer from the gradient vanishing problem.The residual network pushes the model's depth to extremely deep by proposing an identity mapping and addresses the gradient back-propagation well.We analyze the impact of training strategy and model architecture of a deep ConvNet.Specifically,we propose a random scheme to skip some residual connections which gains better performance.At the same time,we use SparkNet to extract deep features from large-scale data sets,which are all multi-threaded parallel computed.In this thesis,we argue that features at different levels of a ConvNet capture different image properties,which incorporates high-level semantic context and low-level spatial structure from different layers.To highlight the boundary contexts,we combine the deep features with the low-level boundary maps derived from the Laplacian of Gaussian(LoG)edge detector,and incorporate them via the independent component analysis with reference(ICA-R)method.We train the dual network with center-shifted random patches,which are generated to augment the positive training samples.The deep features are embedded into the motion and observation model,which is updated online with both stochastic and periodic schemes to address the drift and occlusion problems.Both qualitative and quantitative evaluations on the three benchmark dataset demonstrate the effectiveness of the proposed methods.Compared with massive classical and effective algorithms,our method achieves more favorable performance than several state-of-the-arts.