Object Extraction and Encoding for Video Monitoring Through Low-Bandwidth Networks

Abstract

Surveillance cameras in smart cities generate a dramatic amount of data each day, that must be transferred through communication channels. Most of this information is useless, just images of static scenarios that must be observed by thousands of monitor people. Even though there are plenty of algorithms that helps to automatically classify this information, the computational effort is still concentrated on transporting (coding/decoding) those images. This work aims to present the first results of applying an efficient software analyzer and encoder for static cameras that struggles to reduce computational effort and data generation. For this purpose, a combination of background segmentation and classification algorithms are used, which separate the moving objects from background and tag them using convolutional neural networks. For storing, a hyper-image organization structure is proposed, used for describing such objects in a minimal way. To have efficient algorithms, GPUs and multi-core processors are used. Several tests were carried out, on long video sequences, showing the amount of bandwidth that is reduced compare to other traditional encoders.