利用候选区域的多模型跟踪算法

发布时间：2019-05-28 03:09

【摘要】：跟踪过程中发生的尺度变化、形变、遮挡是导致模型漂移的重要原因。为了克服模型漂移对鲁棒跟踪的影响,本文提出了一种利用多判别式模型和候选区域的跟踪算法。首先,该算法采用候选区域替代传统的滑动采样,适应跟踪过程中目标的位移和尺度变化。接下来,为了提高目标的表征能力,先用预训练网络提取整幅图片的深度特征,再通过感兴趣区域采样层(ROI pooling layer)快速提取每一个候选区域的深度特征,进一步提高跟踪算法的鲁棒性。最后,运用多模型选择机制进行回撤过去错误的模型更新,并通过调整搜索区域实现对目标的重检测,有效抑制了模型漂移对鲁棒跟踪的影响。实验中,本文算法与相关算法在OTB 2013数据库和UAV 20L数据库上进行了对比。结果表明,本文算法在精确度与成功率上均取得了最优性能,并能有效抑制模型漂移对鲁棒跟踪的影响。
[Abstract]:The scale change, deformation and occlusion in the process of tracking are the important reasons for the drift of the model. In order to overcome the influence of model drift on robust tracking, a tracking algorithm based on multi-discriminant model and candidate region is proposed in this paper. Firstly, the candidate region is used instead of the traditional sliding sampling to adapt to the displacement and scale change of the target in the tracking process. Next, in order to improve the representation ability of the target, the depth features of the whole picture are extracted by the pre-training network, and then the depth features of each candidate region are quickly extracted by the sampling layer (ROI pooling layer) of the region of interest. The robustness of the tracking algorithm is further improved. Finally, the multi-model selection mechanism is used to update the past error model, and the re-detection of the target is realized by adjusting the search area, which effectively suppresses the influence of model drift on robust tracking. In the experiment, the algorithm is compared with the related algorithm on OTB 2013 database and UAV 20L database. The results show that the algorithm achieves optimal performance in both accuracy and success rate, and can effectively suppress the influence of model drift on robust tracking.
【作者单位】：空军工程大学航空航天工程学院;
【基金】：国家自然科学基金项目(61472442) 航空科学基金项目(20155596024)
【分类号】：TP391.41

【相似文献】