Privacy-Preserving Anomaly Detection in Bitstream Video based on Gaussian Mixture Model

Advanced computer vision technologies enable nearly real-time intelligent monitoring of homes by detecting anomalies such as falls and incidents of domestic violence, thereby enhancing home safety. Leveraging affordable home cameras and cloud computing services, this technology offers significant societal benefits. However, privacy concerns present challenges for its deployment in real-world settings. This paper introduces a method for anomaly detection in encrypted bitstream videos. By analyzing the video compression standard, we incorporate new appearance information that captures variations in object sizes alongside existing motion information, enhancing anomaly detection capabilities. We also refine the feature extraction process to reduce the impact of noise. Finally, we employ Gaussian Mixture Model to model the probability distribution of the data for effective anomaly detection. Experimental results demonstrate that our proposed privacy-preserving anomaly detection method achieves a commendable balance between protecting privacy and maintaining high detection performance.