基于标准对冲的深度视觉目标跟踪方法研究

Accurately and robustly tracking target under complex and unknown environments is still a challenging task in the field of computer vision. And it has important research significance and application prospects. The project focuses on the three tracking core modules of “target precision modeling”, “model effective combination” and “model intelligent update” to research object tacking by three aspects. Firstly, aiming at accurate target modeling, enhancing image under various environments such as illumination or blur is used to extracted the rich local features, and pre-train the network with existing data resources is trained to enhance the network generalization ability to accurately represent the target. And a stack self-encoding neural network combined with particle filtering and standard hedging is proposed to track object. Secondly, for the effective combination of models, a convolutional neural network tracking method based on fusion correlation filtering and standard hedging is proposed to overcome the boundary effect of correlation filtering end realize the effective combination of the different models. Finally, for the intelligent update of the model, the adaptive online updating and target verification and re-detection mechanism of the target model with interconnected feedback characteristics are proposed to ensure the robustness of the target long-term tracking. The output of this project can provide scientific technical principles and solutions for intelligent target tracking in any unknown environment.

实现未知环境下精准、长时、鲁棒的目标跟踪是计算机视觉领域一项挑战性的任务，具有重要的研究意义和应用前景。本项目围绕 “目标精准建模”、“模型有效联合”和“模型智能更新”三个核心模块，提出：1）针对目标精准建模，提出结合粒子滤波与标准对冲的栈式自编码神经网络目标跟踪方法，在光照、模糊等多变环境下进行增强图像的预处理，提取丰富的目标特征并利用现有数据资源预训练网络，以增强网络泛化能力实现目标精准建模；2）针对模型有效联合，提出融合相关滤波与标准对冲的卷积神经网络目标跟踪方法，根据不同特征分别训练不同的目标相关滤波器构造不同粒子，利用标准对冲克服相关滤波的边界效应，实现模型的有效联合；3）针对模型智能更新，提出具有互联反馈特性的目标模型自适应在线更新和目标验证与重检测机制，保障目标长时跟踪的鲁棒性。所提方法可望为未知环境下的目标跟踪提供新的技术原理和解决方案。

本项目以未知环境下完成精准、长时、鲁棒的目标跟踪这一计算机视觉领域具有挑战性的任务为目标，围绕 “目标精准建模”、“模型有效联合”和“模型智 能更新”三个核心模块展开研究。针对基于粒子滤波和相关滤波框架的标准对冲融合问题，提出了基于标准对冲与特征融合的目标跟踪方法，该方法利用卷积神经网络不同的层的特征以及传统的颜色特征对目标进行精准建模，且在光照、模糊等多变环境下进行增强图像的预处理，确保提取到丰富的目标特征并利用现有数据资源预训练网络，以增强网络泛化能力。针对融合背景感知的深度学习实现网络的训练和优化、特征提取、模型更新为一体的目标协作模型构建问题，提出了提出隐式区域建议网络对目标的形状和尺度信息进行编码学习，首次实现了在无锚跟踪算法中对目标锚盒的隐式嵌入，且实现了自适应目标跟踪。为了提高跟踪效率，利用多层感知机的全局感受野特性，提出了一种简单而有效的基于多层感知器的跟踪（MLPT）有效地对比模板和搜索区域特征，以生成精细特征以及集成信道编码和区块编码处理的信道和空间信息。丰富的实验验证了本项目所提方法的有效性和鲁棒性，相关研究结果可为智能安防监控和智能视觉机器人领域提供理论依据和技术支撑。