Object Recognition And Tracking Based On Structural Sparse Representation

Object recognition and tracking is an important research direction in the field of computer vision.It mainly realizes the categorization judgment and continuous positioning of the targets and has a wide range of applications in military and civil fields,including automatic interpretation of aerospace photos,weapon precise guidance,and other civilian aspects including industrial robot hand eye system and biological intelligent identification.Due to the complexity of application scenarios,the requirements of target recognition and tracking algorithms are also increasing.On the one hand,it has high requirements on the accuracy and robustness of the algorithm,especially when the target is deformed,occluded or interfered by similar objects.On the other hand,it has high requirements on the adaptability of the algorithm in special application scenarios.For example,in the military and security fields,the detection and recognition of targets need to be all-time and all-weather.However,due to the lack of color and texture information in infrared imaging,as well as the lack of large amount of data for feature learning compared with the visual objects,the recognition and tracking for thermal infrared targets are facing greater challenges.This paper focuses on the problem of appearance modeling of visual objects as well as thermal infrared objects in the recognition and tracking task,and conducts in-depth mining and research on the basis of the sparse theory.It adopts a variety of structural sparse coding methods to model the objects in flexible ways.The following five aspects are included in this research:(1)The mathematical basis of the structural sparse theory is used to combine with the practical application of object recognition and tracking tasks.Structural sparse representation model can obtain sparse solutions with special distribution of non-zero elements by controlling the structure function of the dictionary and exploring the potential relationship of the represented samples.Such a coding method can introduce more prior information of the data and improve the flexibility and reliability of the representation model.This is also the theoretical basis of the whole paper.(2)A data-driven adaptive structural local sparse model is proposed,which is effective to deal with the local appearance changes of the target in the recognition task,especially when the image noise has a non-Gaussian distribution.By analyzing the corruption degree of each local block,the image noise can be adaptively reduced.Then the de-noised image is further used in the representation process.The experimental results show that the model can be effectively used to recognize and capture the target before tracking.(3)A locally weighted reverse sparse model based on multi-task learning is proposed for target representation.Meanwhile,the similarity of each candidate image is measured according to the structural characteristics of the sparse coefficients.To eliminate the interference of abnormal templates,this thesis adopts the multi-task learning to represent multiple templates jointly.Moreover,the spatio-temporal information of the template samples is also introduced in the modeling process,making the reliable templates play a more important role.The tracking results show that this method has a significant improvement on the reverse sparse representation model and better tracking performance compared with other state-of-the-art trackers.(4)A mask sparse representation model is proposed for thermal infrared target tracking,which combines high-level semantic features of the target with sparse representation.This new model is able to find specific parts of the target which are distinguishable to its surrounding background.These parts are defined as reliables parts,whose representation results are utilized to guide the representation of other parts of the target.The proposed method is able to weaken the appearance change of the target and background clutter in the thermal infrared target tracking process.The results show that the proposed method has an excellent tracking performance on the VOT-TIR2016 dataset.(5)A bottom-up and top-down integration framework is proposed to achieve online object tracking.Robust online tracking entails integrating shot-term memory based trackers and long-term memory based trackers in an elegant framework to handle structural and appearance variations of unknown objects in an online manner.The bottom-up part employs the short-term memory based trackers,which can generate regions of interest in the input image with a data-driven way.For the top-down part,a graph guided sparse representation model is designed,which learns the high-level perceptual relationship of interested regions,and then estimates the likelihood values.The proposed tracking framework makes full use of the advantages of these two kinds of trackers and achieves good performance in online object tracking.In this paper,4 algorithms are proposed to solve the target modeling problems in object recognition and tracking from different perspectives,which complete the visual tasks from target acquisition to continuous positioning.Furthermore,the proposed algorithms are validated and compared on several open datasets.The experimental results reveal the superiority of the proposed methods.