基于多尺度时空混合专家增强的视频异常检测方法

MSMOE: Enhancing video anomaly detection via multi-scale spatio-temporal mixture of experts

摘要:

针对传统视频异常检测技术中异常数据稀缺导致的对视频特征理解不足，多异常环境下适应性较差的问题，提出了基于多尺度时空混合专家增强的视频异常检测方法；具体来说，使用I3D提取视频特征，从空间和时间两个维度下对视频特征进行多尺度增强，提升对视频特征的理解能力；利用混合专家模块对增强的多尺度特征进行异常得分回归，增强多类型异常检测的鲁棒性；同时，利用多尺度时空对应关系建立回归损失函数，进一步提高了检测精度；在UCF-Crime数据集和ShanghaiTech数据集上的实验结果表明，所提出方法的视频的异常检测准确率有明显提升，多异常场景下检测效果较为稳定,满足实际需求。

Abstract:

Aiming at the problems of insufficient understanding of video features and poor adaptability in multi-anomaly environments caused by the scarcity of anomaly data in traditional video anomaly detection techniques, a video anomaly detection method based on multiscale spa-tio-temporal hybrid expert enhancement is proposed; specifically, the video features are ex-tracted using the I3D, and multiscale enhancement is performed under the spatial and tem-poral dimensions, to improve the comprehension of video features; the enhanced multi-scale features are regressed on the anomaly scores using the hybrid expert module to regress the anomaly scores on the enhanced multiscale features to enhance the robustness of multi-type anomaly detection; at the same time, a regression loss function is established using the multiscale spatio-temporal correspondence, which further improves the detection accuracy; experimental results on the UCF-Crime dataset and ShanghaiTech dataset show that the accuracy of anomaly detection of the videos with the proposed method has been significantly The detection effect is more stable under multiple anomaly scenarios, which meets the practical requirements.