基于行人属性自适应学习的跨摄像机多目标跟踪研究

In recent years, the society has paid increasing attention to safety precaution against violence and terrorist attacks, criminal investigation and so on. Therefore, the technology of multi-target tracking across cameras with high precision has been widely applied in many fields. Compared with the low-level visual features, the middle-level semantic attributes are more robust to the complex appearance variance of pedestrians. In this work, to solve the problem of the lack of pedestrian attribute data and the difference between the domain in the target videos and the domain in the training dataset, we use the spatio-temporal contextual constraints in videos to collect the training samples online. We design the convolution neural network architecture for pedestrian attributes, as well as the loss function and the training method, to adaptively learn more discriminative pedestrian attributes in the target videos. The work is to improve the discriminative ability of pedestrian attributes, and overcome the inter-class changes of different pedestrians and the intra-class appearance variance of the same person due to the factors such as attitude, perspective, scale, occlusion and so on. Furthermore, we build a probability graphic model with the extracted pedestrian attributes to describe the relationship between object detections and object identities. We iteratively solve the maximum conditional probability of the graphic model and update the parameters of the model. We present to associate object detections from local to global to form object trajectories. The work not only provides a new method of modeling and analysis, but also provides more technical supports for higher-level tasks such as video understanding and analysis.

近年来，随着社会对暴恐事件防范、刑事犯罪侦察等安全防范的需求日益增加，高精度的跨摄像机多目标跟踪技术具有广阔的应用前景。与低层视觉特征对比，中层的语义属性对行人的复杂外观变化具有更强的鲁棒性。针对行人属性训练数据缺少、待跟踪视频与训练数据集存在行人属性分布差异的问题，本研究利用视频时空上下文约束信息在线收集训练样本，设计针对行人属性的卷积神经网络结构、损失函数及训练学习方法，提高行人属性判别能力，克服多摄像机视频中不同行人的类间变化，以及同一个行人由于姿态、视角、尺度、遮挡等因素引起的类内变化。进一步融合行人属性建立描述检测响应与目标身份关系的概率图模型，迭代求解概率图模型的最大条件概率和更新模型参数，从局部到全局自动关联检测响应形成目标跟踪轨迹，为多目标数据关联问题提供新的建模和分析方法，为更高层次的视频理解与分析提供更多技术支持。

随着社会对暴恐事件防范、刑事犯罪侦察等安全防范的需求日益增加，高精度的跨摄像机多目标跟踪技术具有广阔的应用前景。与低层视觉特征对比，中层的语义属性对行人的复杂外观变化具有更强的鲁棒性。针对行人属性训练数据缺少、待跟踪视频与训练数据集存在行人属性分布差异的问题，本研究利用视频时空上下文约束信息在线收集训练样本，设计针对行人属性的卷积神经网络结构、损失函数及训练学习方法，提高行人属性判别能力，克服多摄像机视频中不同行人的类间变化，以及同一个行人由于姿态、视角、尺度、遮挡等因素引起的类内变化。进一步融合行人属性建立描述检测响应与目标身份关系的概率图模型，迭代求解概率图模型的最大条件概率和更新模型参数，从局部到全局自动关联检测响应形成目标跟踪轨迹，为多目标数据关联问题提供新的建模和分析方法，为更高层次的视频理解与分析提供更多技术支持。本项目在执行期间共发表期刊论文4篇，包括1篇中科院一区、计算机-人工智能TOP期刊、CCF-A类期刊，以及2篇中科院二区期刊，1篇中文核心期刊，1篇CCF-B类会议论文；申请了3篇专利已受理；参加了两次国际会议，其中一次会议交流在小组作口头报告；协助培养了3名硕士生，都已顺利毕业，另外还协助指导2名硕士、3名博士生（目前在读未毕业）；指导研究生参加教育部举办的第五届中国研究生智慧城市技术与创意设计大赛中完成跨摄像机行人检索任务，并获国家级二等奖；建立了一套跨摄像机多目标跟踪数据库，完成了一套跨摄像机多目标跟踪原型系统。