基于深度稀疏学习的鲁棒视觉跟踪

视觉跟踪中,高效鲁棒的特征表达是复杂环境下影响跟踪性能的重要因素。提出一种深度稀疏神经网络模型,在提取更加本质抽象特征的同时,避免了复杂费时的模型预训练过程。对单一正样本进行数据扩充,解决了在线跟踪时正负样本不平衡的问题,提高了模型稳定性。利用密集采样搜索算法,生成局部置信图,克服了采样粒子漂移现象。为进一步提高模型的鲁棒性,还分别提出了相应的模型参数更新和搜索区域更新策略。大量实验结果表明:与当前主流跟踪算法相比,该算法对于复杂环境下的跟踪问题具有良好的鲁棒性,有效地抑制了跟踪漂移,且具有较快的跟踪速率。

Robust visual tracking based on deep sparse learning

In visual tracking,the efficient and robust feature representation plays an important role in tracking performance in complicated environment.Therefore,a deep sparse neural network model which can extract more intrinsic and abstract features was proposed.Meanwhile,the complex and time-consuming pretraining process was avoided by using this model.During online tracking,the method of data augmentation was employed in the single positive sample to balance the quantities of positive and negative samples,which can improve the stability of the model.The local confidence maps were generated through dense sampling search to overcome the phenomenon of sampling particle drift.In order to improve the robustness of the model,several corresponding strategies of updating model parameters and searching area are proposed respectively.Extensive experimental results indicate the effectiveness and robustness of the proposed algorithm in challenging environment compared with state-of-the-art tracking algorithms.The problem of tracking drift is alleviated significantly and the tracking speed is fast.