The emerging technology of Wi-Fi sensing promises significant benefits for a variety of embedded and edge systems. Using only the radio signals already generated by Wi-Fi interfaces under normal operation, Wi-Fi sensing can theoretically enable an embedded device to detect the presence of humans, estimate their motion, approximate their location, and even sense gestures and subtle movements, such as breathing and heartbeats.
Smart home, entertainment, security, and safety systems can all benefit from this ability. For example, a small sensor in a car could detect the presence of back-seat passengers—soon to be a requirement in new passenger vehicles. It can even detect a child breathing under a blanket as it does not require line of sight. Or an inexpensive wireless monitor in a home could detect in a room or through walls when a person falls—a lifesaver in home-care situations.
Figure 1 Wi-Fi Sensing can be performed on any Wi-Fi-enabled device with the right balance of power consumption and processing performance. Source: Synaptics
Until recently, such sensing could only be done with a passive RF receiver relying on the processing capability of a nearby Wi-Fi access point. Now, it can be done on every Wi-Fi-enabled end device. This article explores how designers can get from theory to shipped product.
How it works
The elegance of Wi-Fi sensing is that it uses what’s already there: the RF signals that Wi-Fi devices use to communicate. In principle, a Wi-Fi receiving device could detect changes in those RF signals as it receives them and, from the changes, infer the presence, motion, and location of a human in the area around the receiver.
Early attempts to do this used the Wi-Fi interface’s receive signal strength indicator (RSSI), a number produced by the interface periodically to indicate the average received signal strength. In much the same way that a passive infrared motion detector interprets a change in IR intensity as motion near its sensor, these Wi-Fi sensors interpret a change in RSSI value as the appearance or motion of an object near the receiver.
For instance, a person could block the signal by stepping between the receiver and the access point’s transmitter, or a passing person could alter the multipath mix arriving at the receiver.
RSSI is unstable in the real world, even when no one is nearby. It can be challenging to separate the influences of noise, transmitter gain changes, and many other sources from the actual appearance of a person.
This has led researchers to move to a richer, more frequently updated, and more stable data stream. With the advent of multiple antennas and many subcarrier frequencies, transmitters and receivers need far more information than just RSSI to optimize antenna use and subcarrier allocation. Their solution is to take advantage of channel state information (CSI) in the 802.11n standard. This should be available from any compliant receiver, though the accuracy may vary.
Figure 2 Wi-Fi system-on-chips (SoCs) can analyze CSI for subtle changes in the channel through which the signal is propagating to detect presence, motion, and gestures. Source: Synaptics
CSI is reported by the receiver every time a subcarrier is activated. It is essentially a matrix of complex numbers, each element conveying magnitude and phase for one combination of transmit and receive antennas. A three-transmit-antenna, two-receive-antenna channel would be a 3 x 2 array. The receiver generates a new matrix for each subcarrier activation. So, in total, the receiver maintains a matrix for each active subcarrier.
The CSI captures far more information than the RSSI, including attenuation and phase shift for each path and frequency. In principle, all this data contains a wealth of information about the environment around the transmitter and receiver. In practice, technical papers have reported accurate inference of human test subjects’ presence, location, motion, and gestures by analyzing changes in the CSI.
Capturing presence data
Any compliant Wi-Fi interface should produce the CSI data stream. That part is easy. However, it is the job of the sensor system to process the data and make inferences from it. This process is generally divided into three stages, following the conventions developed for video image processing: data preparation, feature extraction, and classification.
The first challenge is data preparation. While the CSI is far more stable than the RSSI, it’s still noisy, mainly due to interference from nearby transmitters. The trick is to remove the noise without smoothing away the sometimes-subtle changes in magnitude or phase that the next stage will depend upon to extract features. But how to do this depends on the extraction algorithms and, ultimately, the classification algorithms and what is being sensed.
Some preparation algorithms may simply lump the CSI data into time bins, toss out outliers, and look for changes in amplitude. Others may attempt to extract and amplify elusive changes in phase relationships across the subcarriers. So, data preparation can be anything from a simple time-series filter to a demanding statistical algorithm.
Analysis and inference
The next stage in the pipeline will analyze the cleansed data streams to extract features. This process is analogous—up to a point—to feature extraction in vision processing. In practice, it is quite different. Vision processing may, for instance, use simple numerical calculations on pixels to identify edges and surfaces in an image and then infer that a surface surrounded by edges is an object.
But Wi-Fi sensors are not working with images. They are getting streams of magnitude and phase data that are not related in any obvious way to the shapes of objects in the room. Wi-Fi sensors must extract features that are not images of objects but are instead anomalies in the data streams that are both persistent and correlated enough to indicate a significant change in the environment.
As a result, the extraction algorithms will not simply manipulate pixels but will instead perform complex statistical analysis. The output of the extraction stage will be a simplified representation of the CSI data, showing only anomalies that the algorithms determine to be significant features of the data.
The final stage in the pipeline is classification. This is where the Wi-Fi sensor attempts to interpret the anomaly reported by the extraction stage. Interpretation may be a simple binary decision: is there a person in the room now? Is the person standing or sitting? Are they falling?
Or it may be a more quantitative evaluation: where is the person? What is their velocity vector? Or it may be an almost qualitative judgment: is the person making a recognizable gesture? Are they breathing?
The nature of the decision will determine the classification algorithm. Usually, there is no obvious, predictable connection between a person standing in the room and the resulting shift in CSI data. So, developers must collect actual CSI data from test cases and then construct statistical models or reference templates, often called fingerprints. The classifier can then use these models or templates to best match the feature from the extractor and the known situations.
Another approach is machine learning (ML). Developers can feed extracted features and correct classifications of those features into a support vector machine or a deep-learning network, training the model to classify the abstract patterns of features correctly. Recent papers have suggested that this may be the most powerful way forward for classification, with reported accuracies from 90 to 100% on some classification problems.
Wi-Fi sensing implementation
Implementing the front-end of an embedded Wi-Fi sensing device is straightforward. All that’s required is an 802.11n-compliant interface to provide accurate CSI data. The back-end is more challenging as it requires a trade-off between power consumption and capability.
For the data preparation stage, simple filtering may be within the range of a small CPU core. After all, a small matrix arrives only when a subcarrier is activated. But more sophisticated, statistical algorithms will call for a low-power DSP core. The statistical techniques for feature extraction are also likely to need the power and efficiency of the DSP.
Classification is another matter. All reported approaches are easily implemented in the cloud, but that is of little help for an isolated embedded sensor or even an edge device that must limit its upstream bandwidth to conserve energy.
Looking at the trajectory of algorithms, from fingerprint matching to hidden Markov models to support vector machines and deep-learning networks, the trend suggests that future systems will increasingly depend on low-power deep-learning inference accelerator cores. Thus, the Wi-Fi sensing system-on-chip (SoC) may well include a CPU, a DSP, and an inference accelerator.
However, as this architecture becomes more apparent, we see an irony. Wi-Fi sensing’s advantage over other sensing techniques is its elegant conceptual simplicity. But something else becomes clear as we unveil the true complexity of turning the twinkling shifts in CSI into accurate inferences.
Bringing a successful Wi-Fi sensing device to market will require a close partnership with an SoC developer with the right low-power IP, design experience, and intimate knowledge of the algorithms—present and emerging. Choosing a development partner may be one of the most important of the many decisions developers must make.
Ananda Roy is senior product line manager for wireless connectivity at Synaptics.
Related Content
- Going the Distance With Wi-Fi HaLow
- Exploring the superior capabilities of Wi-Fi 7 over Wi-Fi 6
- Paving the Way to Ambient-Powered Wireless Connectivity
- WIRELESS/RF: SenSiFi WiFi sensor module promises battery life of 3+ years
- Blocking 6-GHz Wi-Fi Is Costing Consumers Money and Quality Experiences
The post How Wi-Fi sensing simplifies presence detection appeared first on EDN.