He Lijun, Qiao Xiaoya, Wen Shuai, Li Fan
Department of Information and Communication Engineering, School of Electronic and Information Engineering, Xi'an Jiaotong University, Xi'an 710049, China.
Sensors (Basel). 2018 Feb 13;18(2):572. doi: 10.3390/s18020572.
Object tracking is an important research direction in computer vision and is widely used in video surveillance, security monitoring, video analysis and other fields. Conventional tracking algorithms perform poorly in specific scenes, such as a target with fast motion and occlusion. The candidate samples may lose the true target due to its fast motion. Moreover, the appearance of the target may change with movement. In this paper, we propose an object tracking algorithm based on motion consistency. In the state transition model, candidate samples are obtained by the target state, which is predicted according to the temporal correlation. In the appearance model, we define the position factor to represent the different importance of candidate samples in different positions using the double Gaussian probability model. The candidate sample with highest likelihood is selected as the tracking result by combining the holistic and local responses with the position factor. Moreover, an adaptive template updating scheme is proposed to adapt to the target's appearance changes, especially those caused by fast motion. The experimental results on a 2013 benchmark dataset demonstrate that the proposed algorithm performs better in scenes with fast motion and partial or full occlusion compared to the state-of-the-art algorithms.
目标跟踪是计算机视觉中的一个重要研究方向，广泛应用于视频监控、安全监测、视频分析等领域。传统的跟踪算法在特定场景下表现不佳，比如目标快速运动和被遮挡的场景。候选样本可能会因其快速运动而丢失真实目标。此外，目标的外观可能会随着运动而改变。在本文中，我们提出了一种基于运动一致性的目标跟踪算法。在状态转移模型中，候选样本通过目标状态获得，该目标状态是根据时间相关性预测得到的。在外观模型中，我们使用双高斯概率模型定义位置因子，以表示候选样本在不同位置的不同重要性。通过将整体和局部响应与位置因子相结合，选择可能性最高的候选样本作为跟踪结果。此外，还提出了一种自适应模板更新方案，以适应目标的外观变化，尤其是由快速运动引起的变化。在2013年基准数据集上的实验结果表明，与现有最先进算法相比，所提出的算法在快速运动以及部分或完全遮挡的场景中表现更好。
