{"title":"Research on Privacy Protection Video Behavior Recognition Method Based on Improved SlowFast","authors":"Yunxue Shao, Min Li, Lingfeng Wang","doi":"10.1002/cpe.70225","DOIUrl":null,"url":null,"abstract":"<div>\n \n <p>Surveillance cameras in public areas have had a positive impact on reducing violence, but have also raised concerns about privacy invasion. With these factors in mind, the Bullying10K neuromorphic dataset utilizes dynamic vision sensor (DVS) cameras to detect instances of violent behavior while protecting individual privacy. This dataset collects a variety of complex, fast violent actions from real-life scenarios as well as potentially occluded individuals. In this paper, for the characteristics of this dataset, the representative SlowFast neural network in behavior recognition is selected for research and improvement. First, a linear interpolation is applied to the Slow pathway to account for the absence of background influence in the Bullying10K dataset. Second, for the complex and fast action features and noise effects in the dataset, the interframe difference method is applied to the input of Fast pathway, which can effectively amplify the real dynamic signals and recognize the motion information in the video. Finally, it is difficult to prevent leakage of nonfacial information, such as gait data, against this dataset. The spatiotemporal attention fusion module (STAFM) is introduced, which not only better protects the privacy of nonfacial information but also improves the security and accuracy of the model when dealing with sensitive data, as well as enhances the generalization ability of the model. Experiments on Bullying10K show that the improved SlowFast exhibits significant advantages, including higher recognition accuracy, better protection of personal privacy, and better generalization capabilities. In addition, this paper is also validated on the UCF101 dataset, and the experimental results demonstrate the generalization of the improved method. The code of this paper is open-sourced at: https://github.com/MinL0128/STAFM-SlowFast.</p>\n </div>","PeriodicalId":55214,"journal":{"name":"Concurrency and Computation-Practice & Experience","volume":"37 21-22","pages":""},"PeriodicalIF":1.5000,"publicationDate":"2025-08-21","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"Concurrency and Computation-Practice & Experience","FirstCategoryId":"94","ListUrlMain":"https://onlinelibrary.wiley.com/doi/10.1002/cpe.70225","RegionNum":4,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q3","JCRName":"COMPUTER SCIENCE, SOFTWARE ENGINEERING","Score":null,"Total":0}
引用次数: 0
Abstract
Surveillance cameras in public areas have had a positive impact on reducing violence, but have also raised concerns about privacy invasion. With these factors in mind, the Bullying10K neuromorphic dataset utilizes dynamic vision sensor (DVS) cameras to detect instances of violent behavior while protecting individual privacy. This dataset collects a variety of complex, fast violent actions from real-life scenarios as well as potentially occluded individuals. In this paper, for the characteristics of this dataset, the representative SlowFast neural network in behavior recognition is selected for research and improvement. First, a linear interpolation is applied to the Slow pathway to account for the absence of background influence in the Bullying10K dataset. Second, for the complex and fast action features and noise effects in the dataset, the interframe difference method is applied to the input of Fast pathway, which can effectively amplify the real dynamic signals and recognize the motion information in the video. Finally, it is difficult to prevent leakage of nonfacial information, such as gait data, against this dataset. The spatiotemporal attention fusion module (STAFM) is introduced, which not only better protects the privacy of nonfacial information but also improves the security and accuracy of the model when dealing with sensitive data, as well as enhances the generalization ability of the model. Experiments on Bullying10K show that the improved SlowFast exhibits significant advantages, including higher recognition accuracy, better protection of personal privacy, and better generalization capabilities. In addition, this paper is also validated on the UCF101 dataset, and the experimental results demonstrate the generalization of the improved method. The code of this paper is open-sourced at: https://github.com/MinL0128/STAFM-SlowFast.
期刊介绍:
Concurrency and Computation: Practice and Experience (CCPE) publishes high-quality, original research papers, and authoritative research review papers, in the overlapping fields of:
Parallel and distributed computing;
High-performance computing;
Computational and data science;
Artificial intelligence and machine learning;
Big data applications, algorithms, and systems;
Network science;
Ontologies and semantics;
Security and privacy;
Cloud/edge/fog computing;
Green computing; and
Quantum computing.