本文从3个方面对原始压缩跟踪算法进行改进, 以提高其在复杂场景下的鲁棒性和准确性。首先, 提出一种结合特征在线选择的压缩跟踪算法, 通过计算相邻两帧同维特征所服从的高斯分布曲线的Hellinger距离来度量特征的置信水平, 从特征池中选择置信水平较高的特征, 并融合特征的置信水平构造贝叶斯分类器。然后, 在压缩跟踪框架下引入协方差矩阵以增强算法对目标的表达能力, 把Haar-like特征和协方差矩阵相结合构建目标模型, 取最大响应值所对应的候选样本作为跟踪结果。最后, 优化分类器参数的更新方式, 根据目标模板与跟踪结果的相似度来自适应更新分类器参数。改进算法的平均跟踪成功率比原算法提高了25%, 平均跟踪精度比原算法提高了22%。相比于原始压缩跟踪算法, 本文算法具有更高的跟踪鲁棒性和准确性。

To improve robustness and accuracy of original compressive tracking algorithm in complex scenes, improvement measures were carried out from three aspects in this paper. First, a compressive tracking algorithm combining with online feature selection was introduced,then the confidence level of the feature was measured by calculating Hellinger distance in the Gaussian distribution curve to which the same dimensional features of two adjacent frames.By selecting the feature with higher confidence level from the feature pool to construct the Bayesian classifier through integrating the confidence level; then, covariance matrix was introduced under the compressive tracking framework to enhance expressive ability of the algorithm towards the target, subsequently, combined Haar-like with covariance matrix to create the target model and selected the candidate sample corresponding to the maximum response value as the tracking results; finally, updated mode of the classifier parameters was optimized: adaptively updating of the classifier parameters was implemented in accordance with similarity between the target template and tracking results. It indicates that compared with original algorithm, average success rate of proposed algorithm is improved by 25%, and the average tracking accuracy is improved by 22%. Hence, the algorithm proposed in this paper can achieve higher robustness and accuracy than the original compressive tracking algorithm.