基于卷积网络特征的逆向稀疏建模的目标跟踪

摘要

针对目标跟踪过程中目标受到光照变化、遮挡等因素的影响而导致目标丢失的现象,提出基于卷积网络特征的逆向稀疏建模的目标跟踪算法.将共享权重的卷积神经网络与目标跟踪相结合,利用卷积网络提取出更抽象、更具表达能力的特征,对目标进行重建,改善目标表示的抗变性.为了减少计算量,在粒子滤波跟踪框架下,加入逆向稀疏思想,即只需要对一个正目标模板进行稀疏求解.在模板更新阶段,选择重建残差满足一定阈值的对应特征进行替换.在实验过程中,分别与基于haar、直方图、梯度等传统特征的跟踪算法进行分析对比,结果表明该方法在光照、遮挡、形变方面有较好的性能.%Aiming at the phenomenon that the target is lost due to the illumination change and occlusion during the target tracking process, this paper proposed an object tracking algorithm which was based on convolutional network features for inverse sparse modeling.The convolutional neural network with shared weights was combined with target tracking.Convolutional networks were used to extract more abstract and more expressive features, to reconstruct the target and improve the anti-degeneration of the target representation.In order to reduce the computational complexity, under the framework of particle filter tracking, the idea of reverse sparseness was added, which means that only a positive target template needs to be sparsely solved.In the stage of template updating,the corresponding features whose reconstruction residuals satisfied a certain threshold were selected to be replaced.In the experiment, the tracking algorithms based on haar, histogram, gradient and other traditional features were analyzed and compared respectively. The results showed that this method had better performance in illumination,occlusion and deformation.