复杂场景下基于卷积神经网络目标跟踪算法研究

Visual object tracking is an important problem in computer vision and has many applications including intelligent video surveillance monitoring, augmented reality and human computer interface. Appearance modeling plays a crucial role in the tracking problem. In order to overcome the shortcomings of current appearance models under complex scenes, we are intended to do research on the feature representation, appearance modeling and updating in this proposal. First, we will study object representation based on convolution neural network (CNN) and feature fusion of different layers, which will make the representation more robust under complex scenes. Second, we will study the multi-task appearance modeling based on the CNN model, which learns the classifier and locator simultaneously. Third, online dynamic memory models will be studied, which select representative patches in previous frames and add them into the training set to update the appearance model. And this will make the appearance model more robust and eliminate the model drifting problem.

视觉目标跟踪是计算机视觉领域研究热点之一，在智能监控、人机交互、虚拟现实等领域有广阔的应用前景。表观建模是目标跟踪核心问题。针对现有表观模型在复杂场景下存在的不足，本项目拟从目标特征表示、统计模型建立和模型更新三方面进行研究。主要研究内容包括：1、研究基于卷积神经网络目标特征表示以及不同层特征融合问题，提高目标特征描述在复杂场景下的鲁棒性；2、研究基于卷积神经网络多任务目标表观模型构建，同时学习目标的分类函数和定位函数，使跟踪定位更精确；3、研究在线动态记忆模型，从历史跟踪结果中挖掘出目标外观变化关键帧加入到训练集中更新目标的表观模型，降低模型漂移问题。

物体检测是计算机视觉领域的热点问题，在视频监控、智慧城市领域等领域具有重要的应用价值。尽管目前物体检测取得了很多的进展，但是大部分算法对低分辨率物体检测性能较差。针对目前算法的问题，我们提出了利用对抗生成网络模型对低分辨率的物体进行超分辨率，再进行检测，取得了卓越的检测效果。共发表高水平论文8篇，其中CCF A类期刊1篇，CCF B类期刊2篇，顶级国际会议3篇，其中2篇是CCF A类会议，1篇接受为口头报告论文，接受率仅为2.16%。