Towards the availability of video communication in artificial intelligence-based computer vision systems utilizing a multi-objective function

Yousef O. Sharrab, Izzat Alsmadi, Nabil J. Sarhan

Abstract: Availability is one of the three main goals of information security. This paper contributes to systems’ availability by introducing an optimization model for the adaptation (controlling the capturing, coding, and sending features of the video communication system) of live broadcasting of video to limited and varied network bandwidth and/or limited power sources such as wireless and mobile network cases. We first, analyzed the bitrate-accuracy and bitrate-power characteristics of various video transmission techniques for adapting video communication in Artificial Intelligence-based Systems. To optimize resources for live video streaming, we analyze various video parameter settings for adapting the stream to available resources. We consider the object detection accuracy, the bandwidth, and power consumption requirement. The results showed that setting SNR and spatial video encoding features (with upscaling the frames at the destination) are the best techniques that maximizing the object detection accuracy while minimizing the bandwidth and the consumed energy requirements. In addition, we analyze the effectiveness of combining SNR and spatial video encoding features with upscaling and find that we can increase the performance of the streaming system by combining these two techniques. We presented a multi-objective function for determining the parameter or parameters’ pairing that provides the optimal object detection’s accuracy, power consumption, and bit rate. Results are reported based on more than 15,000 experiments utilizing standard datasets for short video segments and a collected dataset of 300 videos from YouTube. We evaluated results based on the detection index, false-positive index, power consumption, and bandwidth requirements metrics. For a single adaptive parameter, the analysis of the experiment’s outcome demonstrate that the multi-objective function achieves object detection accuracy as high as the best while drastically reducing bandwidth requirements and energy consumption. For multiple adaptive parameters, the analysis of the experiment’s outcome demonstrate the significant benefits of effective pairings (pairs) of adaptive parameters. For example, by combining the signal-to-noise ratio (SNR) with the spatial feature in H.264, a certain optimal parameter setting can be reached where the power consumption can be reduced to $$20\%$$ 20 % , and the bandwidth requirements to $$2\%$$ 2 % from the original, while keeping the Object Detection Accuracy (ODA) within 10% less of the highest ODA.