Indoor Privacy-preserving Action Recognition via Partially Coupled Convolutional Neural Network
Jixin Liu, Leilei Zhang · 2020
With indoor intelligent surveillance gradually integrated into daily life, privacy-preserving of user is facing a major challenge. In order to solve the contradiction between user privacy-preserving and action recognition, we propose a partially coupled convolutional neural network (CNN) model that only uses anonymous videos of low resolution (LR). Firstly, in the pre-process, the optical flow characteristics of low resolution (16×12 pixels) and high resolution (HR) (64×48 pixels) videos are extracted respectively, and continuous multi-frame optical flow superposition is performed to represent the video action. Secondly, super-resolution technology is used to improve the characteristics of LR optical flow. Finally, a partially coupled CNN model is trained. HR data is added to the training, improve the recognition performance of LR data in test. Experiments on two public datasets show that the method we proposed is better than other state-of the-art methods, which has the dual advantages of privacy-preserving and high recognition accuracy.