TLDR: CSIYOLO is a new framework that improves scatter localization in Integrated Sensing and Communication (ISAC) systems. It uses estimated Channel State Information (CSI) from a single base station-user equipment pair to detect and locate environmental scatters. By treating scatter detection as an object detection problem, using an anchor-based deep learning network with optimizations like circular convolution and convolution factorization, and employing a noise injection training strategy, CSIYOLO achieves high accuracy, effectively removes false detections (ghost targets), and is robust to channel estimation errors. It significantly outperforms traditional methods and can be seamlessly integrated into existing communication systems.
Integrated Sensing and Communication (ISAC) is emerging as a cornerstone technology for the next generation of communication systems. It promises to deliver both high-speed data transmission and precise environmental sensing simultaneously. Within ISAC, “scatter sensing” plays a vital role, focusing on identifying and locating objects or “scatters” in the environment. This capability is crucial for advanced applications such as autonomous vehicles, smart cities, and the burgeoning low-altitude economy.
However, existing scatter sensing methods often face significant hurdles. Many rely on specialized waveforms or hardware modifications, which makes them difficult to integrate with current communication infrastructures. Others, while compatible, suffer from limited sensing accuracy, particularly when the number of scatters in the environment is unknown or when channel conditions are noisy.
Introducing CSIYOLO: A Smart Solution for Scatter Sensing
To address these critical challenges, researchers have developed a novel framework called CSIYOLO. This intelligent system is designed to perform scatter localization using only the estimated Channel State Information (CSI) obtained from a standard base station-user equipment pair. Crucially, CSIYOLO operates without requiring any modifications to existing communication waveforms or signal processing pipelines, allowing it to be seamlessly integrated into current systems as a pluggable module.
The CSIYOLO framework operates in two main stages to achieve its high accuracy and robustness:
Stage 1: Anchor-based Scatter Parameter Detection
CSIYOLO approaches scatter parameter extraction as an “object detection” problem, a concept borrowed from computer vision. First, the raw CSI, which is typically in a frequency-antenna domain, is transformed into an “angular-delay domain.” In this new representation, individual scatters appear as distinct “objects” with specific coordinates corresponding to their transmission delays and angles of departure.
Inspired by the highly efficient You Only Look Once (YOLO) architectures, CSIYOLO employs an anchor-based detection method. Instead of trying to define traditional rectangular bounding boxes around scatters, which are not well-suited for this context, the framework uses “anchors” (elements within the channel matrix) to estimate the relative offsets of nearby scatters and their associated confidence levels. This innovative approach allows CSIYOLO to predict continuous scatter coordinates, thereby overcoming the precision limitations often imposed by discrete grid resolutions in traditional methods.
A significant advantage of CSIYOLO is its “adaptive ghost target removal” strategy. Ghost targets are false detections that commonly arise from multi-path interference and noise, severely degrading localization performance. CSIYOLO tackles this by adaptively estimating the confidence of each detected scatter. Detections below a certain confidence threshold are discarded, and remaining detections that are spatially close are merged into a single, more accurate scatter. This process effectively eliminates ghost targets without needing prior knowledge of the actual number of scatters in the environment, a common requirement for many conventional techniques.
Stage 2: CSI-based Scatter Localization
Once the scatter parameters (transmission delay and angle of departure) are accurately extracted from the first stage, CSIYOLO proceeds to determine the precise physical location of each scatter using an “ellipse model.”
The transmission delay of a scatter path defines an ellipse in the 2D plane. The base station and the user equipment act as the two focal points of this ellipse, meaning the scatter must lie somewhere along its perimeter. The angle of departure then provides a crucial second piece of information: it defines a line originating from the base station. The intersection of this line with the ellipse uniquely determines the scatter’s exact position. This elegant geometric approach effectively translates the detected parameters from the angular-delay domain into real-world spatial coordinates.
Optimized Network Design and Training Strategies
To ensure high performance and efficiency, CSIYOLO incorporates several task-oriented optimizations in its network design and training:
- Network Architecture: The framework utilizes an extendable network architecture built upon the U-Net framework, known for its effectiveness in precise localization tasks. It features a “backbone” for extracting rich features, a “neck” for integrating features across different scales, and a “head” that projects these features into the final parameter space. This multi-scale design allows the network to capture scatter features with greater accuracy.
- Circular Convolution: To enhance localization accuracy, particularly for scatters located near the boundaries of the angular domain, CSIYOLO employs “circular convolution.” Unlike conventional zero-padding, which can disrupt the periodic nature of the angular dimension and lead to information loss, circular padding preserves this periodicity, allowing for more accurate feature extraction from boundary regions.
- Convolution Factorization: To reduce computational complexity and resource requirements without sacrificing accuracy, the network uses “convolution factorization.” This technique decomposes standard 2D convolutions into a series of simpler 1D convolutions, significantly cutting down the number of parameters and operations while maintaining the network’s expressive power.
- Noise Injection Training Strategy: To make CSIYOLO robust against inevitable channel estimation errors in real-world scenarios, a “noise injection training strategy” is implemented. During training, Gaussian noise is progressively injected into the channel matrix. This gradual increase in noise level helps the network learn to adapt to varying noise intensities, making it more resilient and stable under imperfect channel conditions.
Also Read:
- Deep Learning Outperforms Differentiable Ray Tracing in Large-Scale Radio Propagation Modeling
- Optimizing Mobile Edge Computing with Fluid Antennas: A Hierarchical AI Approach
Demonstrated Effectiveness and Robustness
Extensive simulations and experiments have validated the superior performance of the CSIYOLO framework. It consistently outperforms conventional methods like MUSIC-FFT and Low-Rank Tensor Decomposition (LTD) in scatter localization accuracy, achieving higher probabilities of detection and F1-scores. Notably, CSIYOLO maintains its high performance even when the number of scatters is unknown, adaptively estimating them and effectively suppressing ghost targets.
The localization Root Mean Square Error (RMSE) achieved by CSIYOLO is remarkably low, often below 0.2 meters, significantly surpassing the grid-resolution limited accuracy of baseline methods. Furthermore, the optimized network design, including circular convolution and convolution factorization, proves highly efficient, offering better performance with substantially fewer parameters compared to generic object detection frameworks like YOLOv5. The noise injection training strategy also ensures CSIYOLO’s robustness, maintaining stable localization performance across various noise levels.
In conclusion, CSIYOLO represents a significant step forward in integrated sensing and communication. By intelligently transforming scatter localization into an object detection problem and applying deep learning with tailored optimizations, it provides an accurate, robust, and highly compatible solution for environmental sensing. The open-source codes for this paper are available here.
