Deep Learning-Based Object Tracking via Compressed Domain Residual Frames

Abstract

The extensive rise of high-definition CCTV camera footage has stimulated both the data compression and the data analysis research fields. The increased awareness of citizens to the vulnerability of their private information, creates a third challenge for the video surveillance community that also has to encompass privacy protection. In this paper, we aim to tackle those needs by proposing a deep learning-based object tracking solution via compressed domain residual frames. The goal is to be able to provide a public and privacy-friendly image representation for data analysis. In this work, we explore a scenario where the tracking is achieved directly on a restricted part of the information extracted from the compressed domain. We utilize exclusively the residual frames already generated by the video compression codec to train and test our network. This very compact representation also acts as an information filter, which limits the amount of private information leakage in a video stream. We manage to show that using residual frames for deep learning-based object tracking can be just as effective as using classical decoded frames. More precisely, the use of residual frames is particularly beneficial in simple video surveillance scenarios with non-overlapping and continuous traffic.