基于随机掩码的视频异常检测

Video anomaly detection based on random masked Transformer

随着监控摄像机的广泛应用,视频异常检测的技术显得至关重要.基于只有正常训练视频的假设,提出了一种独特的利用随机遮掩技术进行异常检测的方法.该方法包括对视频序列中的特定视频片段进行遮掩,以促使时间Transformer能够有效地提取特征.此外,还设计了一个时间Transformer block和一个空间Transformer block,以实现时空特征的提取.基于空间和时间Transformer,将异常定义为预测帧与真实帧之间存在显著差异的异常.为了更有效地进行运动估计,同时提出了基于时间维度梯度的计算方法,相较于基于光流的方法更具优势.公共数据集上的实验结果表明,随机屏蔽Transformer方法在视频异常检测方面具有显著的有效性.

With the increasing popularity of various types of surveillance cameras,video anomaly event detection has become increasingly important.In this paper,we introduce a novel method for anomaly detection that utilizes a random masked Transformer.This approach involves masking a portion of the video patches in the video sequence to be able to extract features using a temporal Transformer.Besides,we design a temporal Transformer block and a spatial Transformer block to extract spatial-temporal features.Based on the spatial and temporal Transformers,we define anomalies as those with large differences between the predicted frame and true frame.Furthermore,we propose to use the RGB difference to represent the motion,which is more efficient than the optical flow-based methods.Our experiments on public datasets demonstrate that the proposed method using a random masked Transformer approach can detect anomalies from videos effectively.