自适应权值卷积特征的鲁棒目标跟踪算法

针对传统基于固定权值卷积特征的深度学习跟踪算法在部分视频跟踪失败的问题,提出一种新颖的基于响应图和熵函数的评估各卷积神经网络层跟踪性能的方法. 该方法能根据评估结果自动调整各层的权值系数;同时引入边界框检测机制,当跟踪响应最大值小于给定阈值时,采用滑动窗口采样一定数量的边界框,并对边界框进行评估,生成初始建议边界框;最后在初始建议边界框的基础上进行相关滤波跟踪,并给出模型更新策略。 将文中算法与其他9种算法在OTB-2013视频数据库上进行跟踪仿真,实验结果表明,所提算法具有较高的中心点距离准确率和跟踪成功覆盖率。

Robust object tracking via adaptive weight convolutional features

To solve the tracking failure problem in some videos caused by traditional deep learning tracking algorithms with fixed weight convolutional features, this paper proposes a novel tracking method combing the response map and the entropy function which considers the performance of each layer of convolutional neural networks and automatically adjusts the weight parameters. At the same time, an EdgeBoxes detection scheme is introduced when the maximum value of tracking response is less than a given threshold. A great number of bounding boxes are extracted by a sliding window and are evaluated by the EdgeBoxes detection scheme which generates the original proposal bounding boxes. Finally, the tracking method based on the correlation filter are conducted on the original proposal bounding boxes with the update scheme given. We have tested the proposed algorithm and nine state-of-the-art approaches on OTB-2013 video databases. Experimental results demonstrate that the proposed method has a higher precision and overlap rate.