IEEE Access (Jan 2024)
Unlocking Trust: Advancing Activity Recognition in Video Imagery
Abstract
Anonymous activity recognition is pivotal in addressing privacy concerns amidst the widespread use of facial recognition technologies (FRTs). While FRTs enhance security and efficiency, they raise significant privacy issues. Anonymous activity recognition circumvents these concerns by focusing on identifying and analysing activities without individual identification. It preserves privacy while extracting valuable insights and patterns. This approach ensures a balance between security and privacy in surveillance-heavy environments such as public spaces and workplaces. It detects anomalies and suspicious behaviours without compromising individual identities. Moreover, it promotes fairness by avoiding biases inherent in FRTs, thus mitigating discriminatory outcomes. Here we propose a privacy-preserved activity recognition framework to augment the facial recognition technologies. The goal of this framework is to provide activity recognition of individuals without violating their privacy. Our approach is based on extracting Regions of Interest (ROI) using YOLOv7-based instance segmentation and selective encryption of ROIs using the AES encryption algorithm. Furthermore, we investigate training deep learning models on privacy-preserved video datasets, utilising the previously mentioned privacy protection scheme. We developed and trained a CNN-LSTM based activity recognition model, achieving a classification accuracy of $94~\%$ . The outcomes from training and testing deep learning algorithms on encrypted data illustrate significant classification and detection accuracy, even when dealing with privacy-protected data. Furthermore, we establish the trustworthiness and explainability of our activity recognition model by using Grad-CAM analysis and assessing it against the Trustworthy Artificial Intelligence (ALTAI) assessment list.
Keywords
