一种基于多尺度实例学习的视频事件检测算法

发布时间：2018-07-09 10:56

本文选题：视频事件检测 + 特征　；参考：《电视技术》2017年Z2期

【摘要】：现有的大多数视频事件检测方法首先从视频帧或视频快照中提取特征,然后对特征进行量化和汇集,进而为整个视频生成一个向量表示。最后的汇集步骤虽然简单高效,但是可能丢失时间局部信息,而这些信息对于确定长视频中事件发生的位置具有重要作用,从而削弱了事件检测的准确性。为此,本文首先将每个视频表示为多个"实例",并将其定义为不同时间间隔的视频段。然后,针对每个视频的正实例比例已知和未知两种情况,提出基于多尺度实例学习的检测算法,在将实例标签看成隐藏潜在变量的同时推断出实例标签以及实例尺度的事件检测模型。最后,利用大规模视频事件数据集进行了全面的仿真实验,结果证明了本文算法具有显著的性能提升。此外,算法还可以确定视频中导致正检测的时间段的位置,进而对检测结果做出解释。
[Abstract]:Most of the existing video event detection methods first extract features from video frames or video snapshots and then quantize and aggregate the features to generate a vector representation for the entire video. Although the final aggregation step is simple and efficient, it may lose local information of time, which plays an important role in determining the location of events in long video, thus weakening the accuracy of event detection. In this paper, each video is first represented as multiple "instances" and defined as a video segment with different time intervals. Then, a detection algorithm based on multi-scale case learning is proposed for each video in which the proportion of positive instances is known and the proportion is unknown. When the instance tag is regarded as the hidden potential variable, the instance tag and the event detection model on the instance scale are deduced. Finally, a comprehensive simulation experiment is carried out using large-scale video event data sets, and the results show that the proposed algorithm has a significant performance improvement. In addition, the algorithm can also determine the location of the time period in the video that leads to the positive detection, and then explain the detection results.
【作者单位】：宜宾职业技术学院;四川大学计算机学院;
【基金】：国家自然科学基金项目(61471250)
【分类号】：TP391.41

【相似文献】