自适应位置融合的目标跟踪算法

为提高分层卷积特征目标跟踪算法的实时性和鲁棒性,文中提出了一种基于多个相关滤波器预测位置自适应融合的实时目标跟踪算法。该算法首先提取VGG-19网络的Pool4层卷积特征,通过特征均值比对多通道的特征图进行裁剪,提高算法速度。然后利用不同高斯样本分布训练多个相关滤波分类器,并对所有分类器预测的目标位置进行自适应融合,提高算法对目标姿态变化的鲁棒性;最后采用稀疏模型更新策略,进一步提高算法速度。在OTB100标准数据集上测试本文算法, 实验结果表明,该算法的平均距离精度为86.3%,比原分层卷积特征跟踪算法提高了2.6个百分点,在目标发生遮挡、形变、相似背景干扰等情况时具有很好的鲁棒性;平均跟踪速度为45.2帧/s,是原算法的4倍,实时性能良好。

Adaptive Positions Fusion for Visual Tracking

To improve the real-time and robust performances of hierarchical convolutional features for visual tracking method, the adaptive positions fusion based on multiple correlation filters for visual tracking was proposed. Firstly, features were extracted from Pool4 layer of VGG-19 network. Besides, multi-channel feature maps were pruned by average feature energy ratio to speed up the algorithm. Moreover, it trained several correlation filters with different Gaussian distributions of training samples, and fused all the predicted positions adaptively. Finally, the sparse model update strategy was used to further speed up. The proposed algorithm was evaluated on OTB100 benchmark dataset. The results showed that the average precision was 86.3%, which was 2.6 percentage points higher than the hierarchical convolutional features for visual tracking method. It was robust when there were occlusion、deformation and similarity interference. The average speed was 45.2 frames per second, four times than the original method and had favorable real-time performance.