融合上下文信息及多特征目标跟踪方法研究

Research on Target Tracking Method Based on Context Information and Multi-feature Fusion

针对传统相关滤波跟踪算法中单一特征在复杂环境下出现跟踪失败情况,提出一种融合传统特征、卷积特征及上下文信息的运动跟踪方法.通过固定权重融合目标及上下文信息的方向梯度直方图特征响应图和目标颜色直方图特征响应图,再自适应融合卷积特征响应图以更好地跟踪目标,对目标尺度变化问题采用尺度池方法.在标准测试集(OTB-50)中验证了本文算法,与基于传统特征及上下文信息的算法相比,平均距离精度提高了6. 1%,平均重叠精度提高了4. 7%;与只使用卷积特征的算法相比,平均距离精度虽然只提高了0. 2%,但平均重叠精度提高了7. 2%;与其他主流算法相比,性能也优于其他算法,能够有效提升跟踪目标在尺度变化与背景杂波等情况下的准确性与鲁棒性.

Aiming at the problem that the single feature in the traditional correlation filtering tracking algorithm is not able to track in complex environment,this paper proposes a tracking algorithm based on traditional feature and convolutional feature and context information.First,this paper fuses histogram of oriented gradient feature response of target and context information and target color histogram feature response with fixed weight,and then fuses convolutional feature with adaptive weight to better track the target,and adopts scale pool method to solve the problem of target scale change. The algorithm in this paper is verified on standard test set( OTB-50),comparing with the algorithm based on traditional features and context information,the algorithm improves average distance accuracy by 6. 1%,and improves the average overlap accuracy by 4. 7%;comparing with algorithm which only uses convolutional feature,the algorithm improves the average distance accuracy by 0. 2%,but improves the average overlap accuracy by 7. 2%;comparing with other state-of-the-arts algorithms,the performance of this paper is also better than other algorithms,which can effectively improve the accuracy and robustness of tracking targets in the case of scale change and background clutter and so on.