跨场景大规模人群分析方法研究

Dense crowd analysis is one of the challenging problems in computer vision. The popular dense crowd analysis methods mostly assume that the training and test datasets of their model have the same data distribution. However, in fact, there is significant data distribution difference between the training and test scenes due to the different imaging background, illumination, view point and resolution conditions, etc. As a consequence, the performance of the model would be seriously degraded in the test scene. In this proposal we will study the methods for cross-scene dense crowd analysis based on the domain adaptation and crowd analysis theories to realize the robust application of the analysis model across different scenes. The main contents are as follows: (1) Proposing an attention mechanism based cross-scene dense crowd counting and density estimation co-learning method, which discounts the influence from different scenes and exploits the positive correlation between crowd density estimation and counting. (2) Proposing a feature alignment and multi-model information co-training based cross-scene dense crowd coherent motions detection algorithm, which aligns the feature distribution among scenes and captures the complementarity between different features. (3) Proposing a deep cascade and adversarial nets for cross-scene dense crowd anomaly real-time detection model, which emphasizes the detection efficiency and simultaneously ensures the learned feature of this network is domain-invariance and discriminative.

大规模人群分析是计算机视觉领域热点研究问题之一。已有的大规模人群分析方法大都假设模型的训练和测试集具有相同的数据分布，而现实中由于场景之间背景、光照、视角和分辨率等方面的差异，容易导致训练和测试场景具有明显的数据分布差异，进而使模型在跨场景应用中性能大幅降低。本项目拟结合领域自适应和人群分析理论，研究跨场景的大规模人群分析方法，实现模型的跨场景鲁棒应用。主要研究内容包括：（1）提出基于注意力机制的跨场景大规模人群密度和人数估计深度协同学习方法，削弱场景间干扰因素影响的同时充分利用人群密度和人数估计任务的正相关性；（2）提出基于特征对齐和多模态信息协同训练的跨场景大规模人群一致性运动检测算法，实现场景间特征数据分布的对齐，并挖掘不同特征表达之间的互补性；（3）提出基于深度级联对抗网络的跨场景大规模人群异常行为实时检测模型，注重检测效率的同时保证网络学习的特征表达具有类间判别性和场景自适应性。

已有的人群分析方法主要还是假设模型的训练和测试集具有相同的数据分布，而现实中由于场景之间背景、光照、视角和分辨率等方面的差异，容易导致训练和测试场景具有明显的数据分布差异，进而使模型在跨场景应用中性能大幅降低。本项目重点研究了人群分析中所涉及的人体重识别、人脸活体检测和多姿态人脸识别方法上的研究，通过结合领域自适应、多特征融合、红外图像生成和自适应误差驱动学习等技术，提升了相应模型在跨背景、光照、视角和分辨率等场景差异的识别指标，达到了当时的领先水平，提升了模型的跨场景应用鲁棒性。本项目研究的相关技术申请发明专利4项，在国内外核心期刊和会议上发表文章共计6篇。