Research On Key Technologies Of Real-time Visual Tracking System

Visual tracking is one of the most active topics in the field of computer vision,which aims at describing the trajectory of object of interest in the video sequences.Meanwhile,it is also the basic of image cognition and video analysis.Visual tracking has been widely applied in the scenes such as video surveillance,human-computer interaction,medical diagnosis etc.In particular,with the increasing popularity of smart equipments,such as wearable and embedded devices,real-time requirement has been urgent in visual tracking systems.Real-time visual tracking systems encounter challenges as follows.Firstly,modeling of the objects with complex appearance needs more computer resources,such as computational time and storage space,which will ruin the efficiency of system.Secondly,the appearance of object is affected by illumination,pose and scale vibration,which will weaken the discriminative ability of classifier and degrade the accuracy of tracker.Thirdly,occlusion has been paid more attention in visual tracking.It is still a hard work how to solve the occlusion,especially the partial occlusion and complete occlusion.In this thesis,we start from the framework of tracking-by-detection,study the theory and technologies of visual tracking in depth,and propose some approaches solving these issues of visual tracking system in real time.The main contributions of this thesis contain four points as follows:1.This thesis proposes an algorithm of visual tracking based on random fern and random projection theory,which solves the consumption of resources in real-time visual tracking system caused by the complex appearance model.The algorithm proposes to replace the binary code of random fern with the difference of pixel pairs,which can describe more detail information in image without the incremental of computation burden.As random projection theory suggested,the high dimension of difference fern feature are projected to the low dimension space preserving the topology of origin space.In the low domain space,naive Bayesian classifier is trained as a binary classifier distinguishing the object and its background.The experimental results and theory analysis suggest that visual tracking based on random projected fern can run robustly and fast with low computational and storage resources.2.This thesis proposes a visual tracking approach based on random projected fern and multiple instance learning,which aims at improving the robustness of tracker in the complex situation with illumination variation and object deformation.This approach are based on the framework of multiple instance learning.The positive and negative sets are built after the position of object has been predicted.The pool of weak classifiers associated with random projected fern are calculated from the sets of positive and negative samples.This algorithm choose optimal several weak classifiers from the pool by maximizing the value of log-likelihood function of samples.Then,the strong classifier is consist with the selected weak classifiers as to reduce the influence of noise on the tracker.When the tracker search the position of object in the new frame,we introduce a coarse-to-fine search strategy in order to reduce the time of detection.The experimental results suggest that the proposed method can follow the target more robustly at real-time speed.3.This thesis proposes weak label kernel correlation filter,which aims at improving the speed of correlation filter tracking.We find the kernel matrix and label matrix have the same property,which has higher value near the center and lower value relatively far from the center.Therefore,the label matrix can be relaxed as the kernel matrix passed by a low pass filter,which leads to approximate the support vector coefficient by the low pass filter.This algorithm can reduce the computational burden without losing its precision and effectively enhance the speed of tracker.4.This thesis proposes a correlation filter tracking algorithm based on point trajectory,solving the occlusion problem in visual tracking.Firstly,this algorithm estimates the position of object by kernel correlation filter and chooses the samples of point trajectory around the predicted position.After analysis the motion trajectory of object and context,the point trajectory in each frame are cluster and finally labeled as the object or background.The condition whether the object has been occluded or drifting is based on the distribution state of point trajectory in the bounding box of kernel correlation filter.If the object has not been occluded,the bounding box of object in the current frame is regarded as the template.When the drifting or occlusion occurs,the tracker detects the position of object by comparing the color histogram of candidates and template of object.