Finding regions of interest while video conferencing in bandwidth constrained environment

Video conferencing;multimedia systems;bandwidth constrained networks;region of interest;Density-based spatial clustering of applications with noise (DBSCAN);data mining;video games;bandwidth reduction;object detection;face detection;blurring;Parallelization;Compute unified design architecture (CUDA);NVIDIA;graphics processing unit (GPU);parallel computing

Agarwal, Arpit ; Nahrstedt ; Klara

关键词: Video conferencing;    multimedia systems;    bandwidth constrained networks;    region of interest;    Density-based spatial clustering of applications with noise (DBSCAN);    data mining;    video games;    bandwidth reduction;    object detection;    face detection;    blurring;    Parallelization;    Compute unified design architecture (CUDA);    NVIDIA;    graphics processing unit (GPU);    parallel computing;   

Others  :  https://www.ideals.illinois.edu/bitstream/handle/2142/46828/Arpit_Agarwal.pdf?sequence=1&isAllowed=y

美国|英语

来源: The Illinois Digital Environment for Access to Learning and Scholarship

【 摘 要 】

Video conferencing is a popular application to connect people who are geographically distant from each other. People in different parts of the world use video conferencing, however many of them suffer from low bandwidth availability. Since, video conferencing is a network intensive application, it is not always possible to maintain the level of user satisfaction when the available network bandwidth is low. In such situations, the large video frames do not get through the network properly, causing the video to freeze or become pixelated at the receiver end, which may also put audio and video out of sync from each other. In this work, we have tried to come up with a solution that will reduce the bandwidth requirement for a video conferencing application, hence making it possible to sustain user experience in low-bandwidth situations as well. We identify that for any video frame, there are only some parts of it that are actually important to convey the activity being performed in the video, and the rest of the frame is not very informative to the viewer. We call these parts of the video regions of interest, and have come up with five different techniques to identify such regions automatically from a video frame. The basic idea is to reduce the size of each frame by blurring all the pixels that are not part of the region of interest. All these techniques are computationally intensive, and would increase the latency greatly if performed at each frame. We have parallelized our algorithms so as to reduce the running time of the algorithms making it possible to use the solutions in real time. Experimental results show that our techniques can reduce the bandwidth utilization by 42.4 %, which is a great improvement. Also, we performed user studies which concluded that such partial-blurring based on regions of interest does not affect the quality of video perceived by the viewer, or his understanding of the video.