ADGSC: video anomaly detection algorithm based on graph structure change detection in public places

• Published in: Multimedia tools and applications 2023-10, Vol.82 (25), p.38923-38945
• Main Authors: Jiang, Huaiying, Lyu, Chen, Gao, Yuexiu, Zhuang, Yunliang, Du, Sanjun
• Format: Article
• Language: English
• Subjects: Adaptive algorithms; Algorithms; Anomalies; Background noise; Computer Communication Networks; Computer Science; Data Structures and Information Theory; Deep learning; Fault tolerance; Feature extraction; Machine learning; Multimedia Information Systems; Optimization; Periodic variations; Special Purpose and Application-Based Systems; Training; Video data
• ISSN: 1380-7501; 1573-7721
• DOI: 10.1007/s11042-023-15009-5

In real life, the types of anomalous events are diverse and low-frequency, and the collection and labeling of training data is complex. However, most detection algorithms are based on training data and test data, which are difficult to adapt to various monitoring scenarios. In this paper, we propose a video A nomaly D etection algorithm based on G raph S tructure C hange detection, which we call ADGSC. Firstly, we use key frame technique to pre-process the video and enhance the pseudo-periodicity of the video data. Second, our approach proposes an improved DTW algorithm for pseudo-periodicity estimation, which transforms periodicity estimation into a global matching growth rate optimization problem. Thus, the periodicity calculation no longer requires a priori knowledge or parameter settings and can be automatically computed in practical applications. Then, we stitch the normalized HSV histogram and HOG feature descriptors into feature vectors following the period obtained in the previous step for feature extraction of key frames. Further, a sliding window is used to build a graph model to measure the temporal variation of the video data, and median plot denoising is used to reduce the errors caused by feature extraction and metric methods, reduce background, blur and other noise interference, and improve the detection effect. Finally, we use box-line plots and box-line graphs to make decisions. Since we do not use deep learning methods, the evaluation metrics AUC and ROC applied for deep learning are no longer applicable to this method. Instead, our experiments use precision, recall, and F-value, which are commonly used in anomaly detection, to measure the effectiveness of our method. Experiment results show that our algorithm outperforms other current algorithms with unsupervised, adaptive, fault-tolerant, and real-time performance.