针对目标跟踪中出现的尺度变化、旋转和遮挡等问题, 提出了基于分层卷积特征的自适应目标跟踪算法。利用卷积神经网络提取分层卷积特征, 利用相关滤波算法获取卷积特征响应图, 并通过响应图的加权融合估计目标位置。利用一种边缘检测算法实现尺度自适应跟踪。利用峰旁比判断目标的置信度, 解决遮挡情况下的模板更新问题。利用OTB2013数据集测试所提出的算法, 测试得到所提出算法的整体成功率、精确度分别为0.618, 0.861, 在目标发生尺度变化、旋转和遮挡等情况下, 该算法可以准确、可靠地跟踪目标。

An adaptive object tracking algorithm based on hierarchical convolution features is proposed to solve the problems of variable scale, rotation and occlusion in object tracking. The hierarchical convolution features are extracted using the convolution neural network, the response maps of convolution features are obtained by the correlation filtering algorithm, and the weighted fusion response is employed to estimate the location of object. An edge detection algorithm is used to realize the scale adaptive tracking. Peak-side-ratio is used to judge the object confidence and solve the problem of template updating under occlusion. The proposed algorithm is tested in the OTB2013 database. The overall success rate and the precision of the proposed algorithm is 0.618 and 0.861, respectively. In the case of object scale variation, rotation and occlusion, the proposed algorithm can accurately and reliably track the object.