基于ResNet短期视觉的视频跟踪算法

Video Tracking Algorithm Based on Residual Network and Short-Term Visual Memory

针对复杂环境下跟踪过程中目标容易丢失等问题,提出了一种基于残差网络特征提取的视觉记忆矫正相关滤波算法。首先,通过ResNet不同层提取图像感兴趣的深层特征,只选取具体效果最好的卷积层所提取的特征来训练相关滤波器,得到响应值最大的目标位置。其次在确定位置的基础上进行尺度采样和记忆采样,建立短期记忆尺度金字塔,以此建立尺度相关滤波器,从而实现对目标尺度的准确估计。最后在数据集OTB100中与其它算法进行了比较,实验结果表明,所提算法取得了可观的精确度和跟踪成功率,在能保持一定的实时性的情况下适应光照、尺度变化以及遮挡等复杂环境。

Aiming at the problem that the target is easy to lose in the process of tracking in complex environment,this paper proposes a filtering algorithm for visual memory correction based on residual network feature extraction.Firstly,the deep features of interest were extracted from different layers of ResNet,and only the features extracted from convolution layer with the best concrete effect were selected to train the correlation filter,so as to obtain the target position with the maximum response value.Secondly,scale sampling and memory sampling were carried out on the basis of determining the position,and a short-term memory scale pyramid was established,thereby establishing a scale correlation filter,thus realizing accurate estimation of the target scale.Finally,the algorithm was compared with other algorithms in the data set OTB100.The experimental results show that the proposed algorithm has achieved considerable accuracy and tracking success rate,and can adapt to complex environments such as illumination,scale change and occlusion while maintaining certain real-time.