Research On Object Tracking Algorithms Based On Visible And Infrared Sequences

With the rapid development of computer vision algorithms and the reduction of high-quality sensor prices,object tracking algorithm that uses video sequences to find the location of single or multiple interested targets,has gained extensive attention.However,the complex problems partial occlusions,illumination changes,pose changes,background clutters,abrupt motions and scale changes,as well as the requirements of robustness,adaptability and real-time processing,it remains a great challenge to build an efficient object tracking model with high performance.In this paper,we focus on the construction of single-target object tracking model for visible and infrared sequences,then deeply research the algorithms and construction of feature fusion and apparent model and tracking,and finally propose some new efficient tracking algorithms to address related problems.Experimental results indicate the effectiveness of the proposed algorithms.This work is significant for future development.The main contributions are summarized as follows:1.A novel deep multi-view compressive model for visible and infrared sequences is presented.The ex-tended region proposal network helps to improve the accuracy of candidate box samples,which can automatically changes the size and position of the box.In contrast to traditional tracking approaches that exploit the same or similar structural features for template matching,this approach dynamically manages the new compressive layers to refine the target-recognition performance.This paper presents an attractive multi-sensor fusion method which demonstrates the ability to enhance tracking precision,robustness,and reliability compared with that of single sensor.The integration of multiple features from different sensors with distinct characteristics resolves incorrect merge events caused by the inappropriate feature extracting and classifier for a frame.Long-term trajectories for object tracking are calculated using online support vector machines classifier.This algorithm illustrates favorable performance compared to the state-of-the-art methods on challenging videos.2.A novel object tracking algorithm using convolution neural network is proposed based on a fusionmode aiming to achieve a more powerful representation of both visible and infrared sequences.Although deep learning has been successfully applied to visual tracking,most relevant algorithms track the target by using offline neural network which provides generic expressions from a large number of training images.However,the training process of offline object tracking algorithms is timeconsuming,and the target representation aims at the objects from learning process,but in some case,the algorithm can not be distinguished from the special object.In this chapter,we propose a robust and non-trained online convolutional neural network model that does not require offline large amounts of data and has a robust and efficient object tracking framework.Firstly,the algorithm extracts a set of normalized image patches as the convolution filter,from the target region of visible and infrared images in the first frame.These filters integrate a series of adaptive contextual features around the target.It is beneficial to define object feature map of corresponding targets in subsequent frames.In addition,the algorithm uses the concept of relative tracking to obtain a series of relative convolution filters with a special weight respectively from a series of relative tracking candidate proposals in visible and infrared images.Compared with traditional tracking algorithms based on the classification of background and target,this kind of filter provides the target related information.Finally,this algorithm uses soft-shrinkage algorithm to do the overall denoising part and get the final tracking result.The algorithm provides a lightweight tracking algorithm without any special fusion steps.Compared with the state-of-the-art tracking algorithm,this algorithm also verifies its remarkable tracking performance experimentally.3.The theoretical and experimental comparisons of the above tracking algorithms show that the object tracking algorithm based on the deep multi-view compressive model has some redundancy problems for the tracking part,while the fusion part in the the relative target tracking based on the convolutional neural network produces the imbalance problem.For these two limitations,this section completes experiment of the two improved algorithms.