基于自适应深度表观模型的高效目标跟踪方法研究

With the help of deep learning, which is a powerful appearance modeling approach, the object tracking methods based on Deep Convolutional Neural Network (DCNN) have achieved high accuracy on many benchmark datasets. However, most of the tracking methods based on DCNN have high computational complexity and require a lot of computing resources, which greatly limits their applications. This project aims to study the appearance modeling method based on DCNN for object tracking, and proposes corresponding solutions for some key technical problems in practical applications. We will investigate how to design efficient compression methods for DCNN with sparsity constraints, which takes full advantage of the characteristics of the tracking problem, so that we can improve the speed of on-line extracting the deep appearance model. We are going to propose how to build hierarchical tracking methods that can make full use of shallow and deep features. It can efficiently score the current tracking result and make decisions to select which features for tracking so as to improve the tracking efficiency. We will investigate how to effectively improve the diversity of the deep appearance model and enhance the adaptation of the deep appearance model, so that we can improve the performance of the object tracking. Finally, we are going to build and release a new large-scale object tracking benchmark that is with more refined annotations, and fairer for evaluation so as to better promote the research and development of object tracking.

目标跟踪是对视频进行智能分析理解的一项关键技术，借助于深度学习强大的目标表观表达能力，基于深度卷积网络的目标跟踪方法在很多基准测试中都达到了很高的精度。但是大多数基于深度卷积网络的跟踪方法计算复杂性高，需要很多的计算资源，这就大大限制了其应用范围。本项目旨在研究目标跟踪中基于深度卷积网络的表观建模方法，对其在实际应用中存在的一些关键技术难题，提出相应的解决方案。拟研究如何充分利用目标跟踪问题的特点采用稀疏约束对深度卷积网络进行压缩，从而提高深度表观模型在线提取的速度；拟研究如何充分利用浅层和深层特征的特点构造分级跟踪方法，高效的判断当前跟踪结果的可信度从而决策采用何种特征来跟踪，提高跟踪效率；拟研究如何有效的提高深度表观模型多样性，增强深度表观模型的自适应性，提高目标跟踪的性能；拟研究如何构建一个更大规模、更加精细化、更加公平合理的开源目标跟踪基准平台，以便更好的推动目标跟踪的研究和发展。

目标跟踪是对视频进行智能分析理解的一项关键技术，借助于深度学习强大的目标表观表达能力，基于深度卷积网络的目标跟踪方法在很多基准测试中都达到了很高的精度。 本项目旨在研究目标跟踪中基于深度卷积网络的表观建模方法，对其在实际应用中存在的一些关键技术难题，提出相应的解决方案。研究了如何利用多任务深度双相关滤波器进行跟踪，充分利用了深度网络的多层次特征；研究了如何利用孪生可变形互相关网络进行跟踪，以在线自适应方式将可变形互相关操作引入SiamFC，以实现稳健的视觉跟踪；研究了如何得到高效的嵌入特征来进行跟踪，将可微分且具有封闭形式解的判别模型的求解器集成到卷积神经网络中；研究了如何通过对抗性学习进行基于孪生网络的目标跟踪，使得孪生网络可以更准确地生成响应图，从而提高了跟踪的性能；研究了如何在计算机视觉任务中提高深度学习模型的表征能力以及效率。