Home Video Behavior Recognition based on Convolutional Residual Network with Visual Privacy Protection Mechanism and Temporal Adaptive Module

To address the privacy leakage issue that exists in home surveillance video, and considering that 3D Convolutional Neural Network (3D CNN) simply uses a fixed number of video invariant cores, it does not realize the time difference between videos, this paper proposes a home behavior video recognition algorithm based on computer vision considering privacy protection. The system combines compressed sensing (CS) theory with convolutional residual network of temporal adaptive module (TAM). Firstly, compressed sensing technology is used to make video obtain the required "visual masking domain", and then TAM is inserted into 2D Deep Residual Network (2D ResNet). TAM is composed of two modules cascaded together before and after, which give the temporal information in the video sequence into a mapping based on important sensitive locations and a mapping based on invariant locations, using this system to enable the effective identification of behaviors in compressed video.