面向拥挤监控场景的异常事件检测技术研究

With the widely application of video surveillance, abnormal event detection in crowded scenes, as an important and very difficult problem in video surveillance, received daily raising research interests. The dense moving objects in crowded scenes can induce severe occlusion, causing the extracted features are abound with noise, which will greatly degrade the performance of abnormal event detection. This project targets to establish an effective and robust abnormal event detection model for crowded scenes, based on sparse representation theory. To achieve that, the project will firstly integrate a deep structure with variational approach for computing optical flow, aiming to resolve the problem of large displacement and the problem of object occlusion in crowded scenes, which will benefit performance of abnormal event detection; In dictionary learning, we take into the consideration of context information, non-negative coefficient, and adopt a distance metric robust to the local deformation of noise feature, for obtaining a high representative dictionary. Meanwhile, for reducing the structure redundancy in dictionary, we propose an efficient sparse combination learning framework; To accommodate the variety of scenes, we train a robust and good generalization model, through the online update of dictionary based on event classification; In the end, we blend our abnormal event detection model with an online video synopsis method, for effectively displaying the abnormal events.

随着视频监控的广泛应用，拥挤场景下异常事件检测，作为智能视频监控的重点和难点，正受到越来越多的关注。拥挤场景中运动目标密集，目标间遮挡严重，提取的特征包含大量噪声，导致事件模型建立困难。本项目旨在研究拥挤场景下基于稀疏表示的异常事件检测技术，建立有效、鲁棒的事件检测模型。基于此，本项目将融合深度结构和变分法提取光流特征，以有效解决大尺度运动和拥挤场景中的遮挡带来的光流提取问题，提高事件模型性能；利用上下文信息、非负编码等约束，采用对特征局部变化不敏感的度量距离，以分组词典的训练方式，得到更加有表示力、结构冗余小的视觉词典；为了适应场景的变化，根据检测结果在线更新模型，以得到鲁棒性和泛化能力强的事件模型；最后，采用在线视频浓缩的方式，与检测进行有机融合, 以对异常事件进行有效展现。

随着视频监控的广泛应用，拥挤场景下异常事件检测，作为智能视频监控的重点和难点，正受到越来越多的关注。拥挤场景中运动目标密集，目标间遮挡严重，提取的特征包含大量噪声，导致事件模型建立困难。本项目主要研究内容包括：融合深度结构提取运动等特征，以有效解决拥挤场景中的遮挡带来的特征提取问题，提高事件模型性能；利用上下文信息、非负编码等约束，采用对特征局部变化不敏感的度量距离，以分组词典的训练方式，得到更加有表示力、结构冗余小的视觉词典；采用基于深度的目标检测或者高层语义信息提取算法，用于异常事件分析和检测；采用分层策略，对正常行为和异常行为分别聚类，异常行为保留其原有属性；同时利用增量学习原理，更新模型。本项目在执行期间系统研究了基于稀疏表示的异常事件检测技术，取得了良好的相关成果。共计发表论文10余篇，申请软件著作权3项，学术专著1部，培养硕士研究生多名，晋升副教授1名。对于相关问题的研究揭示了稀疏表示技术在事件检测中的应用方法，对于后续的研究也起到了一定的启发作用。