Visual Tracking Via Structural Discriminative Model

Visual tracking,which is the core technology in the newly proposed "City Brain",has attracted more and more attentions in recent years.Visual tracking tries to provide the estimated locations for one or several target objects,and estimate the corresponding scale information.In the civilian field,visual tracking can help the officers to efficiently retrieve the states and trajectories of target objects based on the monitoring data.In the military field,visual tracking is the crucial component for large amounts of defence-related science and technologies,e.g.,the navigation of drones and missiles.This thesis conducts research on visual tracking in the open environment,and focuses on the applications of structural appearance model in complex scenes.The contributions of this thesis are as follows:First,based on the Bayes theory,this thesis models visual tracking as the process of alternately optimizing the locations and occlusion states for target objects.With the occlusion state as a priori,the algorithm takes the states of target patches in the current and previous frames as latent variables,through which a spatialtemporal topology structure is introduced.Based on the spatialtemporal topology structure,the proposed algorithm exploits the information from the optical flow,and provides more accurate motion estimation than previous methods.With the estimated target location as a priori,this thesis simultaneously exploits the positive and negative samples,and constructs a discriminative occlusion model considering the temporal consistency information.The occlusion model can make better distinctions between occlusion and other challenges(e.g.,deformations),thus yielding better tracking performance.Second,this thesis investigates the spatial information of deep features,and proposes two complementary spatialaware regressions for better representation power.The thesis proposes a ridge regression model with cross-patch similarity,wherein the algorithm can simultaneously learn the regression coefficients and the reliability weight for each subregion.The ridge regression formula also models the relationship between different subregions,making it more robust in handling deformations.The proposed method reformulates the ridge regression model as a neural network,and exploits the back propagation algorithm for end-to-end network training.In addition,a convolution neural network with spatially regularized filters is also proposed.By randomly masking some positions in the filter kernel,the algorithm forces each convolution output to focus on a small and localized region.The thesis finally integrates the responses from two networks(i.e.,the ridge regression network and and the convolution neural network)for target localization.Third,this thesis investigates the problem of structural appearance modeling in the correlation filter formula,and rewrites the correlation filter coefficients as the elementwise product of a base filter and a reliability weight matrix.The base filter aims at learning the discrimination information,while the reliability weight matrix tries to learn the reliability information.By introducing the local response consistency term,the algorithm ensures that the learned base filter equally highlights the entire target object,making the filter coefficients insusceptible to the feature maps.The algorithm jointly optimizes the base filter coefficients and the reliability weight matrix via the ADMM algorithm,enabling the algorithm to adaptively determine the importance of each subregion.This thesis provides an efficient solver for the proposed optimization problem in the Fourier domain.