针对MDNet跟踪算法网络模型中存在的特征稳健性差以及目标背景信息丢失导致跟踪失败的问题,提出一种基于重构特征联合的多域卷积神经网络视觉跟踪算法。基于末端卷积层提取的目标高级特征,使用反卷积操作上采样,获得了包含目标背景信息的重构特征,再通过联合目标高级特征和背景信息的重构特征的方式增强特征的稳健性,达到了有效区分目标和背景的目的,适用于解决跟踪过程中出现的目标遮挡、形变、光照变化等问题。将本文算法分别在OTB50和VOT2015跟踪测试集上进行测试,与MDNet算法相比,跟踪精度提升1.53%,跟踪成功率提升2.03%。

The tracking algorithm always receives an inaccurate object position because of the poor robustness of the features in the multi-domain network tracking (MDNet) algorithm network model and the loss of the target background information. In this study, we propose a multi-domain convolutional neural network visual tracking algorithm based on the combined reconstructed features. This algorithm performs the deconvolution upsampling operation on an advanced object feature to obtain reconstructed features containing the background information. This advanced object feature is extracted using the end convolutional layer and is combined with the reconstructed feature, which can enhance the robustness of the feature and effectively distinguish an object from the background,thereby improving the object tracking accuracy in situations such as object occlusion, illumination change, and object deformation. The proposed algorithm is tested using the OTB50 and VOT2015 tracking test sets. When compared with the MDNet algorithm, the tracking accuracy and tracking success rate of the proposed algorithm are improved by 1.53% and 2.03%, respectively.