基于多级视觉注意力和可分离相关滤波的在线视觉跟踪研究

Online visual tracking plays a key role on intelligence video surveillance, and is also a hot and difficult study topic in computer vision and video processing. With respect to effectively mine deep visual features and robustly construct appearance models in object tracking, this project will conduct researches from two perspectives: multi-level visual attention and separated correlation filter models. For the multi-level visual attention model, we will exploit the deep neural networks to model multiple visual attention schemes (spatial, layer, channel and temporal attentions) in a unified framework. This model will online mine deep visual features to facilitate locating the tracked object in current frames, and therefore will improve the representation ability of object appearance. For the separated correlation filter model, we will propose the separation assumption to divide the original filter into the product of a discriminative filter and a reliable filter, and introduce different regularization terms to depict both discrimination and reliability. The global optimization model will be constructed and then be solved using the alternating direction method of multipliers algorithm, which will lead to a robust object appearance model. Finally, we will combine both multi-level visual attention model and separated correlation filter model to develop robust and accurate tracking algorithms. The extensive experiments will be conducted to verify the proposed new theories and algorithms using large-scale standard benchmarks. The findings of this project will support the theoretical study and technical progress for online visual tracking algorithms.

在线视觉跟踪是智能视频监控的关键环节，是计算机视觉和视频处理领域的研究热点和难点。本项目针对跟踪问题中有效深度特征挖掘和鲁棒外观模型建构问题，从多级视觉注意力模型和可分离相关滤波模型两方面展开研究。针对多级视觉注意力模型，拟在深度神经网络中统一建模空间、层、通道、时间等多种视觉注意力机制，在线挖掘有利于当前时刻定位目标的深度视觉特征，提升目标外观特征表达能力；针对可分离相关滤波模型，拟提出可分离假设将原始滤波器拆分为判决性滤波器和可靠性滤波器的乘积，引入不同形式的正则化项来刻画判决性和可靠性属性，建立整体优化模型并利用交替迭代算法求解，构建鲁棒目标外观模型。随后，结合多级视觉注意力模型和可分离相关滤波模型，提出鲁棒精确的在线视觉跟踪算法。本项目最终提出的新理论、新方法将在大尺度标准测评数据库上进行大量的实验验证，研究成果将对在线视觉跟踪算法的理论研究和技术进步提供依据。

在线视觉跟踪是智能视频监控的关键环节，是计算机视觉和视频处理领域的研究热点和难点。本项目针对跟踪问题中有效深度特征挖掘和鲁棒外观模型建构问题，主要从多级视觉注意力模型和可分离相关滤波模型两个方面系统地研究了在线目标跟踪问题，并将研究经验扩展到目标检测和识别领域。项目提出了联合运动和外观信息跟踪模型、多尺度车标检测模型、特征分组多相机多行人模型等有效的跟踪/检测模型及算法，在精度和速度上均取得了较为突出的进展，在多个标准测评中取得了领先的结果。在本项目资助下，课题组共发表论文8篇（SCI检索6篇），其中包括IEEE TIP、IEEE TITS、ICASSP等国际权威期刊及会议。基于上述部分成果，课题组负责人获国家民委中青年英才等奖励。

内容获取失败，请点击重试