基于CNN的粒子滤波目标跟踪算法研究

针对复杂场景下目标跟踪算法存在的跟踪目标丢失漂移等问题，提出一种粒子滤波框架下基于卷积神经网络（convolutional neural network，CNN）的目标跟踪算法.该算法采用CNN提取跟踪目标的高层语义特征，并引入离线训练方式，提高训练效率以及特征提取的泛化能力；利用粒子滤波算法框架，实现目标运动状态的有效估计；同时采用长时与短时两种更新策略，并引入困难样本挖掘的在线训练方式，以适应目标外观变化与背景干扰等复杂情况.仿真实验结果表明本文算法能有效适应遮挡、光照、剧烈运动等场景.与多个当前的跟踪算法在公开测试样本下进行了结果比较和分析，验证了本算法在解决跟踪目标丢失漂移等问题上的有效性. 

Particle Filter for Object Tracking Based on CNN Feature

In object tracking algorithms, the target drift and missing are easy to happen in complex environments. A robust tracking algorithm was proposed based on convolutional neural network(CNN) under the particle filter framework. In this algorithm, CNN was utilized to get high-level semantic features of targets and a offline pre-train method was used to learn the general feature representation and improve the training efficiency. Using particle filter framework, the algorithm was arranged to get the reliable target motion state. In addition, two kinds of online updating procedure, long-time and short time, were introduced to deal with the target postures change and some other situations. An online hard example mining strategy was also used to improve the online learning efficiency. The simulation results show that the proposed algorithm can be effectively adapt to complex background, such as occluded, illumination changes, pose variations. Furthermore, we evaluate the proposed algorithm on some challenging videos compared with the state-of-the-art algorithms.