针对连续域卷积操作跟踪算法中存在的计算复杂问题和过拟合问题, 提出一种降维卷积因式分解方法和一个紧凑衍生模型。本文选取连续域卷积操作跟踪算法中的部分特征通道、因式分解卷积算子, 并结合高斯-牛顿法和共轭梯度迭代求取降维矩阵和分类器, 以达到降低系统的计算复杂程度的目的。选取服从高斯分布的若干样本, 结合联和概率密度函数及混合高斯分布模型, 构建一个更具代表性的紧凑衍生模型, 既能降低过拟合出现的风险又能保证样本的多样性。通过降低模型更新频率, 减少计算复杂度, 同时降低模型漂移发生的概率。经实验验证, 本文基于四类标准数据集获得的跟踪性能曲线积分比连续域卷积操作跟踪算法最少高出0.8%, 最多高出13%。结果表明, 降维卷积因式分解法和紧凑衍生模型可以有效地提升系统跟踪性能。

The reduced dimensional factorized convolution method and the compact generative model (RFCG) are proposed in order to solve the computational complexity and over-fitting problems in the continuous convolution operators for visual tracking(C-COT). Some feature channels in the C-COT are selected in advance, then the convolution operator is factorized, and the reduced dimension matrix and filter are obtained by combining Gauss-Newton method and conjugate gradient iteration (CG), so as to reduce the computational complexity of the system in this paper. A number of samples subject to the Gaussian distribution are selected and combined with the joint probability density function and the mixed Gaussian distribution model to build a compact generative model with stronger representativeness. The risk of the over-fitting problems can be reduced and the diversity of samples can be guaranteed in this way. Finally, the model updating frequency is reduced to save the calculation time and avoid the drift of the model. Experimental results show that the area under the tracking performance curve (AUC) score of the improved algorithm based on four benchmarks achieves at least 0.8% higher than C-COT, and 13% higher at most. This shows that the RFCG can effectively improve the tracking performance of the system.