A Feature-based Video Transmission Framework for Visual IoT in Fog Computing Systems

The rapid development of the internet of things (IoT)promotes research in smart city and Fog computing. The vast volume of real-time visual data produced from the tremendous end devices in IoT is a big challenge for the network to transmit and for the data center to store. The typical case is the huge volume of visual data produced by the surveillance cameras in a smart city. In this paper, we consider the problem of how to allocate the calculation ability of the Fog node to handle the surveillance data to obtain low delay meanwhile maintain the video quality. To solve this challenge, we attempt to reduce the tremendous video data using deep learning models in the computational Fog node and optimize the transmission function for high efficiency. To reduce data, we extract video feature and keep salient zones with high resolution meanwhile leave the unavoidable distortion in less important areas. To obtain the least transmission delay under the dynamic bandwidth in Fog computing, we model the transmission delay function and solve it by Lagrangian dual decomposition. We make experiments on public dataset Cityscapes and 4G/LTE Bandwidth Log to evaluate our method. The experiment results show that our feature-based image processing method obtains around 68.7% higher average SSIM (structural similarity index)than the traditional HEVC in the salient zones, and our solution reduces the system delay by 71.02 % comparing with the plain transmission method. It proves our solution reduces the video transmission latency meanwhile keeps the SSIM of salient areas in the video.