Study For Visual Tracking Based On Sparse Representation

Visual tracking is a research hot topic in the field of computer vision and is widely used in human-computer interaction,industrial robots,video monitoring,intelligent transportation and a number of other fields.As artificial intelligence has become more developed,and video devices more popular,it is becoming more and more important to do research into visual tracking.Although there are a number of visual tracking technologies emerging at present,individual visual tracking tasks still face significant challenges,such as motion blur,posture change,light change,occlusion,and background clutter,all of which make these tasks difficult.In recent years,those visual tracking algorithms based on sparse representation have achieved remarkable results.Those aforementioned algorithms use templates to model a target and assume that candidates can be sparsely represented by the dictionary composed by templates.In those algorithms,the candidate with the least reconstruction errors is chosen as the tracking result.The key points of visual tracking algorithms based on sparse representations are how to design an effective sparse representation model;and the combination of the sparse representation theory with the knowledge gained from previous visual tracking data and research.In this thesis,the specific research object and innovations are mainly focused on the study of sparse representation models and the utilization of visual tracking prior knowledge:1.A model named Reverse Sparse Representation Tracking based on Trace lasso is proposed in this thesis.This model uses candidates sampled from the current frame as dictionary atoms and uses templates as the target in the representation model.Using such templates as the target can effectively reduce the computational burden and increase the algorithm's tracking speed.This model uses a trace lasso as the regularization of representations and the trace lasso can adaptively select different constraints based upon the similarity of information among candidates.This model can utilize the low-rank property of the matrix of candidates and achieve a representation which is adaptive to the rank of the matrix of candidates.Our tracking algorithm can make a more accurate and stable result by utilizing this adaptive representation.2.In the observation model,the prior information obtained from between the template and the candidate sample is introduced.This prior information reflects the similarity between the template and the candidate sample.By utilizing this prior information,the observation model can modify the representation matrix obtained by the sparse representation model,and thus can effectively reduce the number of tracking failures.