【摘要】 针对相关滤波跟踪框架中深度特征跟踪优势受限和计算存储存在冗余等问题，提出一种多尺度估计和自适应响应融合目标跟踪算法。该算法通过调整高斯标签参数，充分发挥手工特征准确性和深度特征鲁棒性优势，并学习连续域卷积算子融合多分辨率特征；为了减少计算和样本的冗余，通过分解卷积操作对特征进行有监督降维来减少模型参数，采用基于高斯混合模型的动态样本融合，并使用模糊稀疏的模型更新机制提高模型有效性；根据预测质量评估标准，进行自适应响应融合。实验结果表明：该算法在目标发生遮挡、形变和快速运动等多种情况下，具有较好的跟踪有效性。更多还原

【Abstract】 Aiming at the limited advantage of deep feature tracking and redundancy in computing and storage in the relevant filter tracking framework, a multi-scale estimation and adaptive response fusion target tracking algorithm was proposed.In the algorithm, the advantages of accuracy and robustness of manual features and depth features could be full used by adjusting the parameters of Gaussian label, and the continuous domain convolution operators were learned to fuse multi-resolution features; in order to reduce the redundancy of computation and samples, the feature was supervised and reduced by deconvolution operation to reduce the model parameters.The dynamic sample fusion based on Gaussian mixture model was adopted, and the model updating mechanism with sparse paste improved the effectiveness of the model, and the adaptive response fusion was adapted according to the prediction quality evaluation standard.The experimental results show that the algorithm has good tracking efficiency in many cases such as occlusion, deformation and fast motion.更多还原