Research On Object Tracking Algorithm Based On Super-Pixel Local Similarity

Object tracking technology is one of the core tasks in the field of computer vision,such as video surveillance,human-computer interaction,medical diagnosis,intelligent vision navigation and many other fields.Due to the influence of illumination change,shadow,occlusion,object moving mutation and background clutter in the complex scenes,the object tracking technology is greatly challenged.In recent years,although the object tracking technology has made great progress,problems such as low tracking efficiency and inaccuracy of similarity measures still need to be solved.In this research,the key technologies for efficient image object tracking are thoroughly studied and the main research contents and related achievements are as followed:In order to solve the problem of inaccurate weight distribution in Euclidean distance measurement,the method of adaptive local weight learning is proposed.(1)In the visual tracking algorithm of minimum boundary particle optimization and super-pixel local weight learning,positive and negative samples are taken to construct objective functions within a certain range of template center.To obtain local weights,the quadratic programming is used to solve the objective function,when the local similarity of template and candidate samples are calculated,these weights can improve the accuracy of local similarity calculation,then the performance of tracking algorithm is improved.(2)In the super-pixel local weighting metric and the inverse-sparse model,we segment the surrounding target regions into super-pixels in stage of training,and then employ mean-shift on the surrounding target regions super-pixels to obtain clusters.A confidence map is calculated according to the clusters.After that,we combine the template super-pixels with the confidence map to compute the initial super-pixel local weights.In the sparse solution process,due to the local weights,the interference of noise and occlusion are weakened when the similarity of template and candidate sample are measured.Because the traditional tracking algorithm is not high in efficiency,(1)the minimum boundary particle optimization method is proposed to eliminate the candidate samples with smaller observation probability,the similarity matching calculation is carried with the sample with large observation probability to reduce the calculation,and then improve the tracking efficiency;(2)In the visual tracking algorithm using the super-pixel local weighting metric and the inverse-sparse model,the sparse representation model is substituted for the original sparse model to improve the tracking efficiency.Model of target template is represented by candidate samples in inverse-sparse representation model,which can obtain sparse coefficients and reduce the number of sparse coefficients to improve tracking efficiency.Both of the improved algorithms adopt adaptive updating mode instead of each frame update,so that it avoids taking too much time to update the template and local weights,the tracking efficiency is improved.