Privacy Protected Recognition of Activities of Daily Living in Video

This paper proposes a new method to protect the privacy while retaining the ability to accurately recognise the activities of daily living for video-based monitoring in ambient assisted living applications. The proposed method obfuscates the human appearance by modelling the temporal saliency in the monitoring video sequences. It mimics the functionality of neuromorphic cameras and explores the temporal saliency to generate a mask to anonymise the human appearance. Since the anonymising masks encapsulate the temporal saliency with respect to motion in the sequence, they provide a good basis for further utilisation in activity recognition, which is achieved by representing the HOG features on privacy masks. The proposed method has resulted in excellent anonymising performances compared using the cross correlation measures. In terms of activity recognition, the proposed method has resulted in 5.6% and 5.4% improvements of accuracies over other anonymisation methods for Weizmann and DHA datasets, respectively.