In order to solve the problem that the tracking results fall into the local minimum easily and the feature extraction process is too slow due to the utilization of deep learning,we study the robust object tracking algorithm based on the Extreme Learning Machine(ELM)and Discriminative Correlation Filter (DCF).Based on the C-COT algorithm,our method improves its feature extraction way and the optimization method for the confidence map.First,a new feature extraction model is designed by using the multi-layer ELM sparse autoencoders to extract the image features efficiently and replacing the original Convolutional Neural Network(CNN).Second,after the feature extraction model,an Online Sequential Extreme Learning Machine(OS-ELM)is used to construct the target rough location estimation model and the multi-peak detection method is used to get the predicted rough location of the target.Third,the search area of the confidence map is determined according to the preliminary target location to avoid the tracking result getting into the local minimum.Finally,the effectiveness of the proposed algorithm is tested on three visual tracking benchmarks.Experimental results show that the proposed algorithm is robust to occlusion, motion blur and similar targets and has a tracking speed of 12.9times that of the C-COT,effectively improving the tracking accuracy and speed.%为解决目前主流相关滤波跟踪方法中跟踪结果容易陷入局部最优值以及因引入深度学习带来的特征提取过程过慢的问题,提出一种融合极限学习机和相关滤波器的鲁棒性目标跟踪算法.该算法在CCOT算法的基础上对其特征提取方式和置信图的寻优方法进行改善.首先,利用多层稀疏自编码极限学习机技术,设计新的特征提取模型,以代替原来的卷积神经网络,可快速且高效地提取图像特征;其次,在特征提取模型之后,添加在线序列极限学习机,构建目标粗糙位置估计模型,采用多峰检测方法初步求得目标的预测位置;第三,根据初步的目标预测位置确定置信图的搜索区域,避免跟踪结果陷入局部最优值;最后,在3个目标跟踪标准数据集上验证新算法的有效性.实验结果表明,新算法的跟踪速度是C-COT算法的12.9倍,且对遮挡、运动模糊以及相似目标等有较强的鲁棒性,可有效地提高跟踪精度和速度.
展开▼
