基于多高斯相关滤波的实时跟踪算法

针对分层卷积特征目标跟踪算法实时性不足和单分类器对目标表观变化适应能力差的问题，提出多高斯相关滤波器融合的实时目标跟踪算法. 为了加快跟踪算法，提取VGG-19网络的Pool4和Conv5-3层的多通道卷积特征，通过稀疏采样减少卷积特征通道数；为了防止特征减少造成精确度下降，利用不同高斯分布样本训练多个相关滤波器，并对所有分类器预测的目标位置进行自适应加权融合，提高算法对目标姿态变化的鲁棒性；采用稀疏模型更新策略，进一步提高算法速度，使算法具有实时性. 在OTB100标准数据集上对算法进行测试，结果表明，该算法的平均距离精度为86.6%，比原分层卷积特征目标跟踪算法提高了3.5%，在目标发生遮挡、形变、相似背景干扰等复杂情况时具有较好的鲁棒性；平均跟踪速度为43.7帧/s，实时性更好.

Real-time tracking algorithm based on multiple Gaussian-distribution correlation filters

Aiming at the shortage of real-time performance of the hierarchical convolutional features for visual tracking algorithm and the poor adaptability of single classifier to target appearance changes, a real-time visual tracking algorithm based on multiple Gaussian-distribution correlation filters was proposed. Features with high dimensions of convolution channels were extracted from Pool4 and Conv5-3 layers of VGG-19 networks, and the sparse sampling approach was used to reduce the number of convolution channels to speed up the tracking algorithm. In order to prevent the decrease of tracking accuracy caused by the reduction of features, the multiple correlation filters based on different Gaussian-distribution samples were trained and all the predicted target positions were fused by adaptive weights, expecting for the better robustness for target posture changes. The sparse model update strategy was applied to further improve the algorithm’s speed and achieve the real-time performance. Experimental results on OTB100 benchmark dataset showed that the proposed algorithm had an average distance precision of 86.6%, which was 3.5% higher than that of the original hierarchical convolutional features for visual tracking method. The proposed method has better robustness under complex conditions, for example occlusion, deformation, similar background interferences. The average tracking speed was 43.7 frames per second, and it had a better real-time effect.