Research On The Algorithms Of Moving Object Discovering And Tracking Under Visual Scenarios

Discovering a moving object from visual scenarios and tracking it are the low-level and middle-level visual processing tasks,which also are the basis of analyzing the behavior of moving object.The main task of this dissertation is to accurately discover the moving object and calculate its state vector in the initial frame of the video sequence,and then estimate the continuous state vectors in the following video frames.Research on this topic has important theoretical and practical value for intelligent video surveillance,intelligent transportation,weapons system,video compression and other fields,which promotes powerful motivation for researchers.In order to achieve an automatic initialization for moving object tracking algorithm,a moving object discovering algorithm via two-stage precise location is studied.The classical moving object tracking algorithms are based on two types of model,which are generative model and discriminate model.After being studied both models,this dissertation proposed two moving object tracking algorithms based on generative model and a discriminate model tracking algorithm.The contributions of this dissertation as follows:Firstly,this dissertation proposed a moving object discovering algorithm based on color optical flow and visual objectness measures.This algorithm is used to solve the automatic initialization problem for moving object tracking.The moving object discovering is viewed as a two-stage location problem followed the principle from coarseness to fineness.First of all,the visual significant color light flow field is calculated according to the Opponent color theory.After normalizing and threshold processing of the light flow field,the moving object rough position is located.Two objectness measures are defined to judge whether an image region is an object from the visual perspective which are color contrast and edge completeness measures.Then the object bounding box can be precisely marked in the sequence images through the sliding window.The experimental results show that this algorithm can ensure the moving object tracking to be achieved automatically.Secondly,a moving object tracking algorithm is proposed under tensor kernel principle component projection.This algorithm tracking an object based on the nonlinear relationship between object appearance images.Because the unfolding matrices of tensor in different modes which are from the same object lie in the Grassmann manifold,the kernel function of tensor can be defined according to notion of the main angle in the Grassmann manifold.The dimension of samples is reduced by kernel principle component analysis(KPCA)in kernel space.Finally,the dynamic prediction method is integrated into Bayesian filtering framework to predict the object state vector under occlusion environment.Compared with classical algorithms,the proposed algorithm has higher accuracy and robustness under the environments of object changes in pose and scale,and partial occlusion.Thirdly,this dissertation proposed a moving object tracking algorithm based on feature tensor multi-manifold discriminant analysis.This algorithm is used to handle multi-similar-object mutual occlusion scenarios.In the multiple objects scenarios,when an object is occluded by another one,the existing tracking algorithms may fail in the process from occlusion to separation because of the interference of similar objects.The color-gradient-based feature tensor was used to describe object appearance when partial occlusion happens.Through the particle filter template matching algorithm,a prior multi-manifold tensor data set is established.For the purpose of embodying discrimination,the tensor distance was defined to determine the intra-manifold and inter-manifold neighborhood relationship in the multi-manifold space.The multi-manifold discriminate analysis is employed to calculate the multi-linear projection matrices of sub-manifolds.Finally,object state vectors were obtained by combining with Bayesian sequence inference.Meanwhile,the reliability of state vector of the newly obtained object was studied and a multi-manifold updating principle was established.Contrast experimental results show that the proposed algorithm can effectively distinguish the object with similar ones in multi-similar-object mutual occlusion scenarios.Finally,this dissertation proposed a moving object tracking algorithm based on color feature randomly compressed.After being studied the color distribution characteristic of an object in a color video sequence,a new color Haar-like feature is introduced.This feature can be extracted by the random compression method.A weighted naive Bayesian classifier is constructed through the study of feature distribution in samples.Finally,in order to adapt to the changes of object appearance,the classifier parameters are updated in real-time.According to contrast experiments on the open challenged video sequence,the effectiveness of the proposed algorithm is verified.