This work introduces a new approach to localize anomalies in surveillance video. The main novelty is the idea of using a Siamese convolutional neural network to learn a metric between a pair of video patches (spatiotemporal regions of video). The learned metric, which is not specific to the target video, is used to measure the perceptual distance between each video patch in the testing video and the video patches found in normal training video. If a testing video patch is far from all normal video patches, then it must be anomalous. We further generalize the approach from operating on video patches from a fixed grid to arbitrary-sized region proposals. We compare our approaches to previously published algorithms using four evaluation measures and three challenging target benchmark datasets. Experiments show that our approaches either surpass or perform comparably to current state-of-the-art methods while enjoying other favorable properties.

这项工作介绍了一种在监视视频中定位异常的新方法。主要的新颖之处在于使用Siamese卷积神经网络来学习一对视频块（视频的时空区域）之间的度量的想法。并非特定于目标视频的学习量度用于测量测试视频中每个视频补丁与正常训练视频中发现的视频补丁之间的感知距离。如果测试视频补丁距离所有正常视频补丁都不远，那么它一定是异常的。我们进一步概括了从对固定网格的视频补丁进行操作到任意大小的区域建议的方法。我们使用四种评估方法和三个具有挑战性的目标基准数据集将我们的方法与以前发布的算法进行比较。