Application of region-based video surveillance in smart cities using deep learning

Area of Science:

• Computer Vision
• Artificial Intelligence
• Urban Informatics

Background:

• Smart video surveillance is crucial for smart cities, but transmitting high-quality video over low-bandwidth channels is challenging.
• Current video encoding techniques, like high efficiency video coding, have limitations in meeting the demand for high-quality encoding of salient regions (pedestrians, vehicles, roads).

Purpose of the Study:

• To develop an efficient salient region-based video surveillance framework for smart cities.
• To address the limitations of existing encoding methods for critical visual data in urban environments.

Main Methods:

• Integration of a deep learning model to extract salient regions from video frames without information loss.
• Encoding of extracted salient regions to reduce data size while preserving quality.
• Application and testing of the framework in diverse smart city case study environments.

Main Results:

• Achieved a bitrate reduction of 56.92%.
• Maintained a peak signal-to-noise ratio of 5.35 bd.
• Demonstrated high salient region (SR) based segmentation accuracy of 92% and 96% on two benchmark datasets.

Conclusions:

• The proposed framework efficiently encodes salient regions, reducing computational load and improving adaptability for smart city surveillance.
• This approach enhances the quality and efficiency of video data transmission for smart city applications.