多示例深度学习目标跟踪  

Target Tracking Based on Multiple Instance Deep Learning

作　　者：程帅[1] 孙俊喜[2] 曹永刚[1,3] 刘广文[1] 韩广良[3] 

机构地区：[1]长春理工大学电子信息工程学院,长春130022 [2]东北师范大学计算机科学与信息技术学院,长春130117 [3]中国科学院长春光学精密机械与物理研究所,长春130000

出　　处：《电子与信息学报》2015年第12期2906-2912,共7页Journal of Electronics & Information Technology

基　　金：国家自然科学基金(61172111);吉林省科技厅项目(20090512;20100312)~~

摘　　要：为解决多示例跟踪算法中外观模型和运动模型不足导致跟踪精度不高的问题,该文提出多示例深度学习目标跟踪算法。针对原始多示例跟踪算法中采用Haar-like特征不能有效表达图像信息的缺点,利用深度去噪自编码器提取示例图像的有效特征,实现图像信息的本质表达,易于分类器正确分类,提高跟踪精度。针对多示例学习跟踪算法中选取弱特征向量不能更换,难以反映目标自身和外界条件变化的缺点,在选择弱分类器过程中,实时替换判别力最弱的特征以适应目标外观的变化。针对原始多示例跟踪算法中运动模型中仅假设帧间物体运动不会超过某个范围,不能有效反映目标的运动状态的缺点,引入粒子滤波算法对目标进行预测,提高跟踪的准确性。在复杂环境下不同图片序列实验结果表明,与多示例跟踪算法及其他跟踪算法相比,该文算法具有更高跟踪精确度和更好的鲁棒性。To overcome the problem that the deficiency of the appearance model and the motion model often leads to low precision in original Multiple Instance Learning（MIL）, a target tracking algorithm is proposed based on multiple instance deep learning. In original MIL algorithm, the image is not represented effectively by Haar-like feature. To improve the tracking precision, a stacked denoising autoencoder is used to learn image features and express the image representations obtained effectively. Selected feature vector could not be replaced in the original MIL algorithm, which has difficulty reflecting the changes of the target and the background.Thus, some weakest discriminative feature vector is replaced with new randomly generated feature vector when weak classifiers are selected. It introduces new information to the target model and adapts to the dynamic changes of the target.Aiming at the deficiency of using motion model where the location of the target is likely to appear within a radius in original MIL algorithm, the particle filter estimates object＇s location to increase the tracking precision.Compared with the original MIL algorithm and other state-of-the-art trackers in the complex environment, the experiments on variant image sequences show that the proposed algorithm raise the tracking accuracy and the robustness.

关 键 词：目标跟踪 多示例学习 深度学习 弱特征更换 粒子滤波 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]