Visual Privacy-preserving Coding for Video Intelligence Applications: A Compressed Sensing Mechanism via Bee-Eye Bionic Vision

Intelligent video surveillance systems can be used to detect abnormal occurrences in the home environment such as falls or assaults. However, traditional video surveillance systems rely on complete and clear video data to complete behaviour recognition tasks, and they therefore cannot successfully fulfill the requirements of visual privacy protection. The home environment involves considerable privacy-sensitive information. If this information is stolen and misused by others, it could seriously threaten personal privacy. To solve this problem, a bionic coding model of bee eye vision based on compressed sensing, named BCBEV-CS, is proposed in this paper. The model combines the low visual acuity and target perception of bee eyes and introduces the encryption of a random measurement matrix in compressed sensing (CS). On the one hand, the model can filter out low-level visual feature information in images or videos to meet the needs of visual privacy protection, and on the other hand, it can retain certain high-level visual feature information to complete behaviour recognition tasks. In addition, we propose a visual privacy protection (VPP) level evaluation method for the BCBEV-CS model at the human visual level. Finally, to better apply the model to intelligent systems, we build a correlation-based statistical model between visual privacy protection and video behaviour recognition, thereby standardizing the coding range of the BCBEV-CS model.