Wi-Fi CSI-Based region detection for adaptive video compression and localization

Abstract

The growing demand for efficient traffic monitoring and surveillance systems, especially in smart cities, requires continuous video recording and transmission for real-time analysis. Static cameras, such as those monitoring roads or public spaces, often capture footage where large portions of the scene remain unchanged, such as the sky, sidewalks, or areas with no movement. Despite this, high-resolution data is continuously streamed to the cloud, consuming excessive bandwidth and energy. With many of these cameras operating on limited power, it is crucial to find ways to reduce the data load while maintaining relevant information for analytics tasks. Wi-Fi Channel State Information (CSI) based region detection for adaptive video compression helps reduce bandwidth and energy consumption in traffic monitoring systems. By using CSI as a sensing modality, we can predict which parts of a frame contain important motion or activity within the camera’s view. It identifies active regions while compressing static areas, thereby reducing the need to process and transmit the entire frame. This approach conserves bandwidth and extends battery life, making it ideal for low-power cameras while maintaining the quality of essential video data for real-time analysis. Traditional encoding techniques, such as AVC (H.264) [1] and HEVC (H.265)[2], often struggle to effectively capture the salient features of a video because they lack contextual awareness of the content within each frame. Most Region of Interest (RoI) based compression techniques depend on analyzing video frames to identify regions of activity, utilizing methods such as motion estimation [3], background subtraction [4], and deep learning-based approaches [5]. However, these methods tend to be computationally intensive and require direct frame analysis, which renders them less suitable for resource-constrained or energy-sensitive applications. In contrast, we argue that relying on the video itself for RoI identification not only consumes considerable energy but also requires significant time investment. Instead, we propose utilizing CSI generated by the existing wireless radio in the node as a more efficient alternative for identifying RoIs.