基于模型更新与快速重检测的长时目标跟踪

Long-Term Object Tracking Based on Model Updating and Fast Re-Detection

在目标遮挡、光线变化等复杂的跟踪环境下,现有相关滤波跟踪算法无法对目标进行长时间实时稳定跟踪。提出一种基于模型更新与快速重检测的长时跟踪算法。首先,在现有的目标定位与尺度变化的相关滤波跟踪算法基础上搭建长时目标跟踪的框架,提出加入模型监测更新机制,根据最大响应和平均峰响应相关能量值判别进入更新或重检测环节;然后,基于提取描述子特征的重检测方法,将提取特征的比特维数统一降到512进行优化,加快重检测速率。所提算法选取OTB-100中20个有代表性的序列进行测试,成功率评估均值为0.706,精确度评估均值为0.805,平均速度为48.5 frame/s;在自采集的数据集上平均准确率能达到87.65%,能够在尺度变化、遮挡等复杂情况下满足长时跟踪的准确性和实时性要求。

In the complex tracking environments, such as those with target occlusion and illumination change, the existing correlation filtering tracking algorithm can not stably track the target for long periods in real time. A long-term tracking algorithm based on model update and fast re-detection is proposed. The proposed algorithm initially builds a long-term target tracking framework based on the existing correlation filtering tracking algorithm with target location and the scale change and subsequently proposes a model monitoring updating mechanism, which enter into updating or re-detection link according to the correlation energy values of maximum response and average peak response. Then, the bit dimension of the extracted features can be reduced to 512 using the re-detection method based on extracting descriptor features to accelerate the re-detection rate. The algorithm presented in this study selects 20 representative sequences from OTB-100 dataset for testing. The average success rate, average accuracy, and average rate are observed to be 0.706, 0.805, and 48.5 frame/s, respectively. Furthermore, the average accuracy of the self-collected datasets reaches 87.65%, satisfying the accuracy and real-time demands of long-term tracking in complex situations, including scale change and occlusion.