论文部分内容阅读
针对传统卷积网络需要耗费大量离线训练时间和带标签样本的问题,提出一种在线卷积网络模型并应用于视频目标跟踪.首先通过层级K-means算法获取卷积网络不同层的滤波器组,而后通过这些滤波器在线提取目标的层级外观特征图.同时,层间降采样操作被去除以便不损失空间信息.最后得到的特征图通过多通道相关滤波器预测目标的位移变化.实验表明,本文算法可以有效对抗背景纹理、光照变化等噪声干扰,具备较高的跟踪精度.
Aiming at the problem that traditional convolutional networks need to spend a lot of offline training time and labeled samples, an online convolutional network model is proposed and applied to video target tracking.First, the filter banks of different layers of convolutional network are obtained by the hierarchical K-means algorithm , And then through these filters to extract the target’s hierarchical appearance feature map.At the same time, the layer down sampling operation is removed in order not to lose the spatial information.Finally, the feature map through the multi-channel correlation filter to predict the target displacement.Experiments show that, This algorithm can effectively counter the noise of background texture, light changes and other interference, with high tracking accuracy.