长时跟踪中抑制模板漂移关键技术研究

Long-term tracking aims at long-term continuously tracking motion objects under the actual complex application scenes, and object model is an important basis for judging a candidate being a real object. However, most of the existed tracking methods, assuming a single-state model, updated the object model for every frame with a linear weighted style, without judging the correctness of a current tracking result. A few tracking methods judged a current tracking result and updated the object model when the tracking results without an occlusion. When online updating an object model by an inaccurate tracking result or a linear weighted style, an accumulated error can always result in a model drift and a tracking failure. This proposal will research how to restrain the model drift for long-term tracking. Concretely, a confidence judged mechanism will be designed by analyzing the law of a history response-map sequences based on an accumulated first-order derivative function curve. By the mechanism, a current tracking result can be divided into three situations of slight appearance variation or partial occlusion, severe appearance variation, and severe occlusion or disappearance. Aiming at the multi-state characteristic of object appearance by severe appearance variation, a multi-state model mechanism will be constructed introducing the thought of Gaussian Mixture Model. When one state model being matched with a high probability, a model updated mechanism will be realized using a Back Propagation method, for decreasing the accumulated error from updating the object model. The research will form the effective mechanism of restraining model drift for long-term tracking, increasing the robustness of long-term tracking.

长时跟踪旨在实现复杂场景下运动目标的长时间持续跟踪，而目标模板是判别候选目标是否真实目标的重要依据。然而，大多数已有跟踪算法假定目标模板呈现单一状态，且不对跟踪结果正确性进行判别直接对模板逐帧线性加权更新，少量算法进行置信度判别仅在未遮挡情况线性加权更新模板，依据不准确的跟踪结果或线性加权更新方式进行在线更新带来的累积误差导致模板漂移，进而导致现有跟踪算法总是失败。项目将针对长时跟踪中抑制模板漂移关键技术展开研究。具体将基于累积一阶导数函数曲线分析响应图历史序列规律设计置信度判别机制，将当前跟踪结果分为表观微变或局部遮挡、表观剧变、重度遮挡或消失三种情况；针对表观剧变导致目标表观呈现多态性，引入高斯混合模型思想构建多态模板机制；当某一状态模板被高概率匹配时，采用反向传播方法实现模板更新机制，以降低模板更新带来的累积误差。其研究将形成长时跟踪中抑制模板漂移的有效机制，以期提高长时跟踪鲁棒性。

长时跟踪旨在实现复杂场景下运动目标的长时间持续跟踪，而目标模板是判别候选目标是否真实目标的重要依据。项目紧密围绕长时跟踪系统中抑制模板漂移的关键技术展开研究，具体开展了以下三个子问题的研究工作：（1）置信度判别机制。项目组成员深入分析响应图的历史序列规律，提出了五种置信度判别机制：基于累积一阶导数的置信度判别机制、基于最大响应值变化率的置信度判别机制、基于相邻帧相似度的置信度判别方法、基于反向检测方法的置信度判别机制、和相关响应图置信区域的判别方法。（2）目标模板建立机制。构建了两种目标函数：一种引入目标尺度变化信息的目标函数和一种基于自适应空间正则化的畸变抑制目标函数。提出了三种跟踪结果校正方法：基于多边形顶点位置均值的校正方法、基于多边形质心的位置校正方法和基于峰值变化比与状态判别相结合的位置校正方法。（3）目标模板更新机制。在项目提出的响应图置信度判别机制基础上，实现了目标模板的自适应更新策略。还提出了一种基于注意力的目标模板更新机制。基于公开视频数据集与已有算法进行对比的实验结果表明，基于本项目研究成果改进后的目标跟踪算法在长时跟踪过程中取得了更为鲁棒的跟踪效果。其相关研究成果已在国内外重要期刊和会议上发表论文9篇（其中期刊论文6篇、会议论文3篇），申请发明专利3项。