面向强噪声场景的低秩稀疏学习目标跟踪方法研究

Video object tracking usually suffers from kinds of problems, such as object appearance variation and complex environment interference. These problems can seriously impact the accuracy and the stability of the tracking, and lead to tracking drift. Face to complex strong noisy scene, the project researches the algorithms for deformation object tracking. Based on the frequency characteristic and the long memory property of fractional differential, the project researches the dynamic denoising of tracking sequences under the fractional-order variation framework. The processes of dynamic denoising and the later object tracking collaborate in an interlacing manner to improve the robustness of object tracking. Combining sparse and low-rank representation theory, the project builds the appearance model for deformation object, researches the structural characteristic of the training data for penalty the appearance modeling, and designs the weighted constraint strategy for appearance modeling based on visual saliency. For the occlusion problem, the project researches the patch-level appearance modeling method based on local low-rank sparse representation, designs local structural sparse occlusion processing strategy based on the scene depth difference between the object and the occlusion. Finally, face to the object appearance model, in the framework of particle filtering, combining duality theory and projection method, the tracking learning algorithm is designed in a streaming fashion to improve computational efficiency.

视频目标跟踪经常会面临目标外观变化和复杂环境干扰等问题，这将严重影响跟踪的精确性和稳定性，导致目标跟踪漂移。本项目重点开展面向复杂强噪声场景的形变目标跟踪方法的研究工作。基于分数阶微分的频率特性和长记忆性，研究分数阶变分框架下的跟踪序列动态去噪算法，实现动态去噪和后续目标跟踪的交错处理，改善目标跟踪的鲁棒性。结合稀疏表示和低秩近似理论建立形变目标表观模型，研究训练数据结构特征提取方法惩罚表观建模，设计面向视觉显著性差异的表观建模加权约束策略。针对目标遮挡情况，研究基于局域低秩稀疏表示的粒子表观建模方法，并根据目标与遮挡物间具有景深差异的特性，设计面向深度信息的局域结构化稀疏遮挡处理策略。最终，面向目标表观模型，结合对偶理论和投影方法，设计粒子滤波框架下流处理模式的跟踪学习算法，提高目标跟踪的计算效率。

视频目标跟踪技术是计算机视觉领域的重要研究内容，已广泛应用于车载导航、安防监控等领域。但该技术仍面临目标形变、严重遮挡等挑战问题，影响跟踪精度。本课题以强噪声场景下的视频目标跟踪问题为研究对象，利用稀疏表示、低秩表达等数学工具，对目标跟踪的建模和数值算法开展了深入研究。面向强噪声场景，基于分数阶变分理论实现视频序列的动态去噪。针对目标形变问题，表观建模中引入融合罚约束，允许目标表观在个别帧间存在较大差异性，保护跳跃信息。利用变分调整将稀疏系数差分建模在有界变差空间，允许帧间差异存在跳跃不连续性。采用核范数低秩约束目标表观的时域相关性。针对严重遮挡问题，利用混合范数稀疏表示候选粒子的局部斑块，保护候选粒子间及其局部斑块间空间布局结构。引入分数阶变分约束融合多帧间信息相关性。最后，通过反向稀疏表示描述目标表观，用候选粒子稀疏表示目标模板，减少优化问题数目，简化在线目标跟踪计算。