Skip to main content

Overview

Mo Meng Jingcheng (莫梦竟成); Zheng JianKang (郑健康)About 9 min

➢ 认识异常检测

■ 什么是视频异常检测?

视频异常检测(Video Anomaly Detection, VAD)是指在视频序列中自动检测和定位异常事件或行为的任务。异常是指与正常模式显著不同的事件,如交通事故、犯罪行为等。VAD 在安全监控、智能交通系统和公共安全等领域中有广泛的应用。

1. 简单介绍(从异常行为检测到视频异常行为检测)

2. 论文综述

  • [1] 邬开俊等. 视频异常检测技术研究进展 [J]. 计算机科学与探索, 2022。
  • [2] Bharathkumar Ramachandra et al. A survey of single-scene video anomaly detection (TPAMI 2020)。

■ 视频异常检测的挑战

  • 异常定义的模糊性:异常事件的定义通常是主观的,具有多样性,导致检测难度较大。
  • 数据不平衡:正常事件在视频数据中远多于异常事件,这导致训练过程中的数据不平衡问题。
  • 实时性要求:在实际应用中,需要实时检测异常以便及时响应。
  • 场景多样性:视频场景和拍摄角度的多样性增加了检测难度。

■ 视频异常检测方法分类

1. 无监督方法

无监督方法不需要标签数据,仅依赖正常样本来学习正常行为的模式,通过检测偏离这些模式的行为来识别异常。

  • 自编码器:通过重建输入视频帧,若重建误差较高则认为是异常。
  • 记忆增强网络:通过将正常行为存储在记忆模块中,当新输入不匹配这些记忆时,检测为异常。

2. 弱监督方法

弱监督方法使用视频级别的标签进行训练,不需要逐帧标注。

  • 多实例学习:将视频看作由多个实例组成的包,通过学习视频中正常和异常实例的差异来检测异常。
  • 对比学习:通过比较正常和异常数据,学习到区分特征。

■ 经典数据集

1. 无监督数据集

  • UCSD Pedestrian:主要用于行人行为分析,包含行人通道中的异常事件,如骑自行车、滑板等。此数据集通常用于无监督异常检测。
  • ShanghaiTech:较大规模的数据集,包含校园监控视频,涵盖多种异常行为,适用于无监督方法。
  • Avenue:以校园场景为背景,涵盖诸如突然跑步、丢弃物品等异常行为。此数据集适用于无监督异常检测方法。

2. 弱监督数据集

  • UCF-Crime:一个用于弱监督视频异常检测的大规模数据集,涵盖不同场景下的犯罪行为,如偷窃、打架等。视频级标注用于训练和评估模型。
  • XD-Violence:包含各种暴力事件的视频数据集,常用于弱监督方法,帮助模型学习区分正常与异常行为。
  • UCFCrime2Local:UCF-Crime 的子集,包含空间标注,用于在弱监督条件下进行异常事件的精确定位。

■ 评价指标

1. AUC(曲线下面积)

  • 定义:AUC(Area Under the Curve)指的是 ROC 曲线(接收者操作特征曲线)下的面积。ROC 曲线描绘了模型在不同阈值下的假阳性率与真阳性率的关系。
  • 用途:用于衡量分类模型在不同阈值下的整体性能。AUC 值越接近 1,模型的区分能力越强。
  • 优点:在衡量模型的总体表现时,AUC 可以帮助了解模型的稳定性和判别能力,不依赖于具体的阈值。

2. EER(等错误率)

  • 定义:EER(Equal Error Rate)是指假阳性率(False Positive Rate, FPR)和假阴性率(False Negative Rate, FNR)相等时的错误率。
  • 用途:常用于验证模型性能的一个稳定点,尤其是在需要权衡假阳性和假阴性影响时。
  • 优点:EER 可以为不同阈值下的平衡性能提供直接的比较,适用于二分类问题中权重相当的应用场景。

3. F1-score

  • 定义:F1-score 是精确率(Precision)和召回率(Recall)的调和平均数,公式为:

    F1=2×Precision×RecallPrecision+Recall F1 = 2 \times \frac{\text{Precision} \times \text{Recall}}{\text{Precision} + \text{Recall}}

  • 用途:用于在分类问题中权衡模型的精确率和召回率,特别是在类分布不平衡时尤为有效。
  • 优点:F1-score 平衡考虑了精确率和召回率,适合评估那些对假阳性和假阴性影响敏感的模型。
  • 应用场景:适用于异常检测中对误报和漏报都有较高要求的场合。

4. Precision(精确率)

  • 定义:精确率是指模型预测为正样本的实例中实际为正样本的比例。公式为:

    Precision=TPTP+FP \text{Precision} = \frac{\text{TP}}{\text{TP} + \text{FP}}

  • 用途:衡量模型预测结果的准确性,适用于对假阳性率较敏感的应用。

5. Recall(召回率)

  • 定义:召回率是指实际为正样本的实例中被模型正确预测为正样本的比例。公式为:

    Recall=TPTP+FN \text{Recall} = \frac{\text{TP}}{\text{TP} + \text{FN}}

  • 用途:评估模型对正样本的检测能力,适用于对假阴性率较敏感的场景。

每个指标在不同的应用场景中具有不同的重要性。例如,在安全监控中,可能更关注召回率,以确保尽可能少的异常被漏检;而在高误报成本的场景下,精确率则更为重要。

➢ 优秀团队 / 学术大佬

■ 高盛华 - 上海科技大学(视觉与数据智能中心)

  • [1] A Revisit of Sparse Coding Based Anomaly Detection in Stacked RNN Framework (ICCV 2017) 提出了ShanghaiTech数据集。
  • [2] Future Frame Prediction for Anomaly Detection – A New Baseline (CVPR 2018)。
  • [3] Future Frame Prediction for Anomaly Detection (TPAMI 2022)。

■ Radu Ionescu - SecurifAI/University of Bucharest

  • [1] Detecting abnormal events in video using Narrowed Normality Clusters (WACV 2019)。
  • [2] Object-centric Auto-encoders and Dummy Anomalies for Abnormal Event Detection in Video (CVPR 2019)。
  • [3] Anomaly Detection in Video via Self-Supervised and Multi-Task Learning (CVPR 2021)。
  • [4] A Background-Agnostic Framework with Adversarial Training for Abnormal Event Detection in Video (TPAMI 2021)。
  • [5] UBnormal New Benchmark for Supervised Open-Set Video Anomaly Detection (CVPR 2022)。
  • [6] Self-Supervised Predictive Convolutional Attentive Block for Anomaly Detection (CVPR 2022)。

➢ 经典论文推荐(加“👍”)

■ Unsupervised VAD

  • Conference Papers

    • [1] Learning Temporal Regularity in Video Sequences (CVPR 2016)。
    • [2] 👍 Future Frame Prediction for Anomaly Detection – A New Baseline (CVPR 2018)。
    • [3] 👍 Memorizing Normality to Detect Anomaly: Memory-augmented Deep Autoencoder for Unsupervised Anomaly Detection (ICCV 2019) → 首次在视频异常检测中使用记忆模块。
    • [4] 👍 Object-Centric Auto-Encoders and Dummy Anomalies for Abnormal Event Detection (CVPR 2019) → 首次结合对象检测和VAD以实现对象级异常检测。
    • [5] AnoPCN: Video Anomaly Detection via Deep Predictive Coding Network (ACM MM 2019) → 首个混合模型。
    • [6] 👍 Learning Memory-guided Normality for Anomaly Detection (CVPR 2020) → 基于MemAE。
    • [7] Cluster Attention Contrast for Video Anomaly Detection (ACM MM 2020) → 首次应用对比学习。
    • [8] 👍 Anomaly Detection in Video via Self-Supervised and Multi-Task Learning (CVPR 2021) → 对象级。
    • [9] 👍 A Hybrid Video Anomaly Detection Framework via Memory-Augmented Flow Reconstruction and Flow-Guided Frame Prediction (ICCV 2021) → 混合模型。
    • [10] Anomaly Detection in Video Sequence with Appearance-Motion Correspondence (ICCV 2019) → 双流网络。
    • [11] Video Anomaly Detection and Localization via Gaussian Mixture Fully Convolutional Variational Autoencoder → 双流网络。
    • [12] Self-supervised Sparse Representation for Video Anomaly Detection (ECCV 2022) → 首次尝试解决无监督和弱监督VAD。
    • [13] Video Anomaly Detection by Solving Decoupled Spatio-Temporal Jigsaw Puzzles (ECCV 2022)。
  • Journal Papers

    • [1] Video Anomaly Detection with Sparse Coding Inspired Deep Neural Networks (TPAMI 2021)。
    • [2] A Background-Agnostic Framework With Adversarial Training for Abnormal Event Detection in Video (TPAMI 2022)。
    • [3] Influence-Aware Attention Networks for Anomaly Detection in Surveillance Videos (TCSVT 2022)。
    • [4] Bidirectional Spatio-Temporal Feature Learning With Multiscale Evaluation for Video Anomaly Detection (TCSVT 2022)。
    • [5] Anomaly Detection With Bidirectional Consistency in Videos (TNNLS 2022)。
    • [6] Variational Abnormal Behavior Detection With Motion Consistency (TIP 2022)。
    • [7] DoTA: Unsupervised Detection of Traffic Anomaly in Driving Videos (TPAMI 2023)。
    • [8] A Hierarchical Spatio-Temporal Graph Convolutional Neural Network for Anomaly Detection in Videos (TCSVT 2023)。
    • [9] Learnable Locality-Sensitive Hashing for Video Anomaly Detection (TCSVT 2023)。
    • [10] A Kalman Variational Autoencoder Model Assisted by Odometric Clustering for Video Frame Prediction and Anomaly Detection (TIP 2023)。
    • [11] Abnormal Event Detection and Localization via Adversarial Event Prediction (TNNLS 2023)。

■ Weakly Supervised VAD

  • [1] 👍 Real-world Anomaly Detection in Surveillance Videos (CVPR 2018)。codeopen in new window
  • [2] Weakly Supervised Video Anomaly Detection via Center-Guided Discriminative Learning (ICME 2020)。
  • [3] Graph Convolutional Label Noise Cleaner: Train a Plug-And-Play Action Classifier for Anomaly Detection (CVPR 2019)。codeopen in new window
  • [4] Not only Look, but also Listen: Learning Multimodal Violence Detection under Weak Supervision (ECCV 2020)。codeopen in new window
  • [5] CLAWS: Clustering Assisted Weakly Supervised Learning with Normalcy Suppression for Anomalous Event Detection (ECCV 2020)。codeopen in new window
  • [6] Localizing Anomalies From Weakly-Labeled Videos (TIP 2021)。codeopen in new window
  • [7] Learning Normal Dynamics in Videos with Meta Prototype Network (CVPR 2021)。codeopen in new window
  • [8] Weakly-Supervised Spatio-Temporal Anomaly Detection in Surveillance Video (IJCAI 2021)。
  • [9] Weakly-supervised Video Anomaly Detection with Robust Temporal Feature Magnitude Learning (ICCV 2021)。codeopen in new window
  • [10] MIST: Multiple Instance Self-Training Framework for Video Anomaly Detection (CVPR 2021)。codeopen in new window
  • [11] Self-Training Multi-Sequence Learning with Transformer for Weakly Supervised Video Anomaly Detection (AAAI 2022)。codeopen in new window(only readme)
  • [12] Exploiting Completeness and Uncertainty of Pseudo Labels for Weakly Supervised Video Anomaly Detection (CVPR 2022)。codeopen in new window
  • [13] Modality-Aware Contrastive Instance Learning with Self-Distillation for Weakly-Supervised Audio-Visual Violence Detection (ACM MM 2022)。codeopen in new window
  • [14] Decouple and Resolve: Transformer-Based Models for Online Anomaly Detection From Weakly Labeled Videos (TIFS 2023)。
  • [15] Dual Memory Units with Uncertainty Regulation for Weakly Supervised Video Anomaly Detection (AAAI 2023)。codeopen in new window
  • [16] TEVAD: Improved video anomaly detection with captions (CVPR 2023)。codeopen in new window
  • [17] Look Around for Anomalies: Weakly-supervised Anomaly Detection via Context-Motion Relational Learning (CVPR 2023)。
  • [18] Unbiased Multiple Instance Learning for Weakly Supervised Video Anomaly Detection (CVPR 2023)。codeopen in new window
  • [19] Hierarchical Semantic Contrast for Scene-aware Video Anomaly Detection (CVPR 2023)。codeopen in new window
  • [20] MGFN: Magnitude-Contrastive Glance-and-Focus Network for Weakly-Supervised Video Anomaly Detection (AAAI 2023)。
  • [21] Dynamic Erasing Network Based on Multi-Scale Temporal Features for Weakly Supervised Video Anomaly Detection (ArXiv 2023)。codeopen in new window
  • [22] Text Prompt with Normality Guidance for Weakly Supervised Video Anomaly Detection (CVPR 2024)。
  • [23] Prompt-Enhanced Multiple Instance Learning for Weakly Supervised Video Anomaly Detection (CVPR 2024)。codeopen in new window

■ Based on Large Model

  • [1] 👍 Harnessing Large Language Models for Training-free Video Anomaly Detection (CVPR 2024)。codeopen in new window
  • [2] Open-Vocabulary Video Anomaly Detection (CVPR 2024)。
  • [3] 👍 Uncovering What, Why and How: A Comprehensive Benchmark for Causation Understanding of Video Anomaly (CVPR 2024)。codeopen in new window
  • [4] 👍 HAWK: Learning to Understand Open-World Video Anomalies (NeurIPS 2024)。codeopen in new window
  • [5] Video Anomaly Detection and Explanation via Large Language Models (ArXiv 2024)。codeopen in new window
  • [6] VANE-Bench: Video Anomaly Evaluation Benchmark for Conversational LMMs (ArXiv 2024)。codeopen in new window
  • [7] 👍 Holmes-VAD: Towards Unbiased and Explainable Video Anomaly Detection via Multi-modal LLM (ArXiv 2024)。codeopen in new window
  • [8] VadCLIP: Adapting Vision-Language Models for Weakly Supervised Video Anomaly Detection (AAAI 2023)。codeopen in new window

➢ 经典项目

➢ 发现的新的有意思的研究方向 → Explainable Anomaly Detection (EAD) 可解释性异常检测

■ 定义

此任务旨在检测视频中的异常事件并自动生成高层次解释。理解异常事件的原因至关重要,因为所需的响应取决于其性质和严重性。

■ 相关工作

  • [1] Joint Detection and Recounting of Abnormal Events by Learning Deep Generic Knowledge (ICCV 2017)。
  • [2] X-MAN: Explaining multiple sources of anomalies in video (CVPR workshop 2021)。
  • [3] Discrete neural representations for explainable anomaly detection (WACV 2022)。