一种健壮的超像素跟踪算法

在目标跟踪中，传统的超像素跟踪算法在发生遮挡等情况后，会将非目标超像素标记为目标加入到特征空间. 在对候选样本置信度计算中，利用特征空间中最近邻超像素来划定样本中超像素的簇归属会产生错误；而依据的近邻超像素数量过多时，又会造成分类误差的积累. 为解决上述问题，本文提出一种健壮的超像素跟踪算法. 本算法以贝叶斯算法为框架，首先，将前几帧进行超像素切割，提取特征并使用均值漂移聚类算法和基于超像素的外观表示模型进行分类和计算类置信度，放入特征空间中. 其次，根据接下来几帧的平均中心误差确定最佳近邻数目. 最后，在跟踪过程中，对获取帧的指定区域进行超像素切割，提取特征、进行软分类和计算置信度；根据上一帧目标位置进行高斯采样，累加样本内超像素置信度，获得样本置信度；在发生严重遮挡时，不进行滑动窗口更新和外观模型修改，使用当前模型继续跟踪. 与传统的最近邻超像素算法相比，本算法能够有效提升跟踪成功率和降低平均中心误差.

Robust Superpixes Tracking Method

During the object tracking, when occlusion occurs, the traditional superpixel tracking algorithm will add the superpixels of the non-target area into the feature space. In the calculation of the candidate sample confidence, the nearest neighbor superpixel in the feature space is used to delimit the cluster attribution of the superpixels in the sample, and the accumulation of the classification error is caused by the excessive number of neighboring superpixels. To solve the problems above, we propose a robust superpixels tracking method. This algorithm uses Bayesian algorithm as the framework. Firstly, we slice the first few frames into superpixels, extract the feature, use the mean shift clustering algorithm and representation model based on superpixel to classify and calculate the class confidence value, and put the feature into feature space. Secondly, the suitable numbers of neighbors can be found with the mean center error of next few frames. Last but not least, during the tracking process, the superpixel is segmented in the specified area of the acquired frame, to extract the feature. The cluster is confirmed with soft classification and the confidence value is calculated. According to the previous frame target position, the Gaussian sampling is collected. We can obtain the sample confidence value with the accumulation of the confidence value. In case of severe occlusion, the sliding window update and the appearance model modification are not carried out, and we continue to use the current model to track. Compared with the traditional tracking algorithm based on nearest superpixel, the algorithm can effectively improve the tracking success rate and reduce the average center errors.