Discriminant correlation filters (DCF) have proven to perform well in visual tracking. However, the underlying discriminant correlation filter only utilizes the single-resolution features of the image block, which limits the performance of its algorithm. And there is still a problem with the existing DCF, that is, the degradation of the filter over time. The continuous convolution operator filter (C-COT) uses implicit interpolation to combine multi-resolution features to construct learning problems in a continuous domain. This efficiently integrates multi-resolution depth feature maps to solve the problem of targeting on the different resolution features. In order to improve the above problems, based on the continuous convolution operator filter (C-COT), the filter is LASSO regularized and adaptive spatial feature selection can be realized, and the filter maintains a certain sparsity, and secondly A time-aware optimization term is added to the objective function so that the learned filter can learn near its historical value to maintain the global structure of the feature. In order to make the learned filter anti-occlusion, this paper adopts a high confidence update strategy to avoid the model being polluted. The algorithm in this paper is called LASSO regularized time-aware continuous convolution operator filter (LTCOT). The algorithm can not only combine the effective features of multi-resolution, but also make the learning filter fully smooth transition between two frames and have certain sparsity, and has stronger discriminative power and anti-interference ability. The video in the OTBIOO database was used for testing with OPE (one pass evaluation), the area under the success line (AUC), average overlap accuracy (OP), and center pixel error distance (DP). The results show that the LTCOT algorithm’s AUC score on the OTBIOO dataset is 2.4% higher than C-COT, OP is 85.9%, better than C-COT(81.4%), and DP is 90.1%, which is better than 89.3% of C-COT. At the same time, LTCOT’s speed is six times higher than C-COT. The LTCOT algorithm can adapt to the visual tracking of complex situations such as occlusion, rotation and deformation, and has good robustness and accuracy.


    Access

    Check access

    Check availability in my library

    Order at Subito €


    Export, share and cite



    Title :

    Learning LASSO Regularized and Time-Aware Continuous Convolution Operator Correlation Filter for Visual Tracking


    Contributors:
    Zhang, Heng (author) / Cao, Li (author)


    Publication date :

    2019-10-01


    Size :

    1217647 byte




    Type of media :

    Conference paper


    Type of material :

    Electronic Resource


    Language :

    English



    Articulated tracking with manifold regularized particle filter

    Gonczarek, A. | British Library Online Contents | 2016


    Deep Correlation Filter Tracking With Shepherded Instance-Aware Proposals

    Liang, Yanjie / Wu, Qiangqiang / Liu, Yi et al. | IEEE | 2022


    CONVOLUTION OPERATOR SELECTION

    MCGILL STEPHEN G / DREWS PAUL / ROSMAN GUY | European Patent Office | 2021

    Free access

    Convolution operator selection

    MCGILL STEPHEN G / DREWS PAUL / ROSMAN GUY | European Patent Office | 2025

    Free access