军事知识服务平台
基于异常引导学习的弱监督视频异常检测
摘要 视频异常检测(video anomaly detection,VAD)旨在识别视频序列中偏离正常行为模式的异常片段或事件.随着公共安全与智能监控需求的不断增长,该领域得到了快速发展.近年来,弱监督视频异常检测(weakly supervised video anomaly detection,WSVAD)因其较低的标注成本而受到广泛关注.然而,由于多实例学习(multiple instance learning,MIL)框架面临数据集复杂度高和视频噪声干扰严重等挑战,其检测性能仍受到较大限制,这些问题显著阻碍了模型对异常模式的有效学习.为解决上述问题,本文提出了一种基于异常引导学习的弱监督视频异常检测方法.具体而言,该方法利用异常相关表征对模型学习过程中的偏差进行校正.首先,通过关键片段学习模块和异常分类模块分别提取关键片段表征与异常类别表征;其次,设计异常引导注意力机制对两类表征进行嵌入与融合,从而增强模型对异常语义信息的感知能力.在 ShanghaiTech 和 UCF-Crime 数据集上的大量实验结果表明,所提出的方法能够显著提升异常检测精度,并取得具有竞争力的性能表现.
Abstract:
Video anomaly detection(VAD)aims to identify snippets or events that deviate from normal behavior in video sequences.With the growing demand for public safety and surveillance,this field has developed rapidly.In recent years,weakly supervised VAD(WSVAD)has attracted increasing attention.However,its performance remains limited due to the challenges of multiple instance learning(MIL),including high dataset complexity and substantial video noise,which severely hinder model learning.To address these issues,we propose a WSVAD method guided by anomaly representations.Specifically,our method leverages anomaly-related representations to correct biases in the learning process.It extracts key snippet and anomaly category representations through the key snippet learning module and anomaly classification module,respectively.These two types of representations are then embedded via an anomaly-guided attention mechanism.Extensive experiments on the ShanghaiTech and UCF-Crime datasets demonstrate that the proposed method significantly improves detection accuracy and achieves competitive performance.