基于相关滤波和卷积神经网络的目标跟踪算法

在目标跟踪系统中,获得目标的良好表征是确定目标跟踪性能的关键,因此提出一种基于相关滤波和卷积神经网络的目标跟踪算法;该算法首先在各视频场景内预先选定可清晰区分目标外观的参考区域块用以构造训练样本,并构建了两路不完全对称但权值共享的卷积神经网络;该卷积神经网络使得参考区域外目标的输出特征尽可能与参考区域内目标的输出特征相似,以便于获得参考区域内目标的良好表征,并在其中一路加入了相关滤波模块,实现了卷积网络与相关滤波的结合;实验结果验证了该算法的可行性。

Target Tracking Algorithm Based on Correlated Filters and Convolutional Neural Network

In target tracking system, the key to obtaining good characterization is to determine target tracking performance, therefore, this paper proposes a target tracking algorithm about the correlated filters and convolutional neural networks. This algorithm firstly pre-selects reference blocks which can distinguish the target appearance in each video scene to construct the training samples and then build the two-way convolutional neural network which is not completely symmetric and which shares the weights. This convolutional neural network makes the target output characteristics outside the reference area as similar as possible to the target output characteristics in the reference area so as to get good characterization of the target in the reference area. The correlated filter module is added into a way to realize the combination of convolutional network and the correlated filter. Experimental results verified the feasibility of the algorithm.