SHARP: Environment and Person Independent Activity Recognition with Commodity IEEE 802.11 Access Points

1 INTRODUCTION

• Processing Wi-Fi sensing measurements into features for target recognition under unknown environmental conditions remains an open issue.The paper situates this gap alongside the emerging use of deployed Wi-Fi systems for indoor sensing.Source: §1 · p. 1
• SHARP uses phase-cleaned CFR data to extract micro-Doppler traces, which a neural network uses to classify human activities.Doppler shift reveals moving scatterer velocities and is not affected by static objects such as furniture and walls.Source: §1 · p. 2
• The study covers dynamic activities; recognizing static poses is out of scope because sub-6 GHz features cannot reliably distinguish a still person from an empty room.The authors note that a static person introduces no micro-Doppler effect.Source: §1 · p. 2
• 96% average accuracy in the most challenging person-, day-, and environment-shift scenario; SHARP also exceeds 95% accuracy across seven activities.The 96% result covers four activities; the seven-activity result adds standing, arm gymnastics, and sitting down/standing up.Source: §1 · p. 2
• SHARP is evaluated with commercial IEEE 802.11ac devices and can be integrated without modification with OFDM-based IEEE 802.11n/ax/be/bf standards.The authors release the CFR dataset and code for replicability and benchmarking.Source: §1 · p. 2

2.1 CSI based human activity recognition

• Earlier CSI-based HAR studies often did not target generalization to unseen environments and subjects; the authors identify that gap as important for Wi-Fi sensing systems.They report that one approach's accuracy fell from 99% to 80% on other datasets.Source: §2.1 · p. 3

2.2 Exploitation of the CFR phase

• SHARP corrects CFR phase offsets using the strongest path within each sample as a reference, avoiding a separate reference measurement.The literature review notes that other approaches use reference antennas, cables, or time-averaged signals for phase correction.Source: §2.2 · p. 3

2.3 Doppler based applications

• Prior Wi-Fi Doppler systems included tracking and exercise recognition, while used reference-antenna sanitization and an SVM but did not assess unseen training scenarios.SHARP is described as outperforming and recognizing activity in at least 95.99% of cases in unknown environments.Source: §2.3 · p. 3
• SHARP's design rationale is that Doppler can separate reflections from static structures and moving targets, assuming the environment is mostly static.The paper presents extracting Doppler from 802.11ac COTS devices and pairing it with dedicated processing and learning architectures as its contributions.Source: §2.3 · pp. 3–4

3 CSI DATA PROCESSING

• CSI is the environment-dependent CFR matrix, estimated per packet; when existing network traffic cannot be captured, the setup uses one transmitting access point and a monitor.The channel parameters are estimated from known preamble symbols.Source: §3 · p. 4

3.1 Phase sanitization

• A sparse delay-grid reconstruction identifies paths; multiplying each subchannel by the strongest path’s conjugate removes phase terms shared across paths, including the offset.The method uses a strongest-path reference because the phase offset is common to the paths.Source: §3.1 · pp. 4–5

3.2 Doppler trace computation

• Static poses are out of scope because sub-6 GHz, off-the-shelf Wi-Fi features cannot reliably distinguish them from an empty room.The paper focuses on dynamic human activities.Source: §3.2 · p. 5
• Assuming the strongest path is static, an STFT over N-packet windows extracts Doppler vectors; vp cos αp proxies motion-driven path changes without separate speed and angle estimates.The window length is selected so attenuation, velocities, and angles can be considered constant within each window.Source: §3.2 · pp. 6–7
• Empty-room Doppler traces show power mainly at zero velocity, while a moving person spreads power across bins, reflecting human-related multipath changes.Fig. 3 compares the empty room with sitting, walking, running, and jumping.Source: §3.2 · p. 7, Figure 3 · p. 7
• SHARP is trained in one scenario and tested without retraining as the day, person, and/or environment change.The evaluation examines generalization beyond the training setup.Source: §3.2 · p. 7

4 LEARNING ARCHITECTURE FOR HAR

• SHARP computes Doppler traces from channel estimates across receiving antennas, then applies a neural-network prediction chain separately to each antenna stream.The activity label can be updated at every new CSI sample using a sliding window.Source: §4 · p. 7
• Combining antenna data at the classifier input did not yield an environment-invariant classifier because outcomes depended on antenna ordering and the person’s location.Source: §4 · pp. 7–8

4.1 Activity classifier for the single antenna system

• The single-antenna classifier processes Doppler traces through parallel, differently sized kernels and outputs one of five activity classes.The proposed network has 128,535 parameters; it reduces the feature maps before a five-neuron dense layer.Source: §4.1 · p. 8, Figure 4 · p. 8

4.1.1 On the selection of the learning architecture

• The Inception module targets small- and large-scale movement patterns; deeper or adaptive multiscale alternatives were excluded as more complex without significant expected gains for this task.The authors prioritize a lightweight approach suitable for low-cost devices.Source: §4.1.1 · pp. 8–9

4.2 Decision fusion for the multiple antenna system

• Each antenna produces an independent activity prediction, combined by majority agreement or, without a clear winner, by summing activity vectors.The final label is the class with the highest score in the summed vector.Source: §4.2 · p. 9

5.1 Nexmon extraction tool

• The monitor router uses Nexmon to extract channel information from 242 data sub-channels per transmit-receive antenna pair at 80 MHz bandwidth.The experiments use one transmitting antenna and four receiving antennas at the monitor.Source: §5.1 · p. 9

5.2 Dataset acquisition and organization

• Two active Wi-Fi terminals provide signals of opportunity, while a passive monitor router records channel estimates as one person performs activities indoors.The campaigns cover a bedroom, living room, and university laboratory with three volunteers; the setup is intended to reflect sensing over existing Wi-Fi traffic.Source: §5.2 · pp. 9–10
• Set S1 trains the model; S2 tests a different day, S3 changes the person, and S7 changes the room, day, and person relative to training.The seven sets vary environment-day-person conditions and include measurement station, activity area, and direct-path status.Source: §5.2 · p. 10, Table 1 · p. 10
• Each campaign records 120 seconds per activity plus 120 seconds in an empty room, totaling nearly 120 minutes of CSI data across the four monitor antennas.Campaigns span several months in 2020 and January 2022.Source: §5.2 · p. 10

6.1 Pre-processing steps

• SHARP’s preprocessing normalizes CFR amplitude, sanitizes phase, constructs 245-component CFR vectors, and stacks 340 Doppler vectors into roughly two seconds of input.Each Doppler vector uses 31 sanitized CFR samples, is zero-padded to 100 points, and is thresholded to remove contributions below 12 dB.Source: §6.1 · p. 11

6.2 SHARP training and performance assessment

• 95.99% average HAR accuracy is achieved when both the person and room differ from training, with walking most often confused with running.Across the reported tests, mean recognition accuracy exceeds 95%; it approaches 100% when the monitor environment and location match training, regardless of day or person, and is around 97% when the direct path is blocked.Source: §6.2 · p. 11, Figure 7 · p. 11

6.3 On the selection of the phase sanitization approach

• Above 95% accuracy is retained with ref. main path across S2–S7, whereas one-antenna sanitization degrades when the person or environment changes.For sitting followed by running, the proposed method produces a clearer Doppler spectrogram; the one-antenna reference can remove relevant environmental-change information, while the static-component reference retains noisy contributions.Source: §6.3 · p. 12, Figure 8 · p. 12, Table 4 · p. 12

6.4 On the selection of the fusion approach

• Data fusion is less effective than decision fusion, especially with blocked line of sight or when both the test environment and subject differ from training.With data fusion, antenna inputs are combined before classification, giving each antenna similar influence; decision fusion combines per-antenna outputs and can discount a discordant prediction.Source: §6.4 · pp. 12–13, Table 5 · p. 13

6.5 On the impact of the number of monitor antennas

• Accuracy increases as the number of monitor antennas rises, while remaining above 80% even with a single antenna.The authors associate multiple antennas’ spatial diversity with greater robustness to changing environmental conditions.Source: §6.5 · p. 13, Figure 9 · p. 14

6.6 HAR algorithms performance comparison

• Competing HAR systems match SHARP on S1 but degrade in other scenarios, while SHARP generalizes across days, environments, and people.The comparisons include DeepSense and EI using CSI amplitude and MatNet-eCSI using phase and amplitude; the authors attribute SHARP’s generalization to Doppler features and its Inception-based classifier.Source: §6.6 · p. 13, Figure 9 · p. 14

6.7 SHARP for smart-living applications

• SHARP recognizes eight activities with high accuracy in the reported single-subject, line-of-sight setting, but the three added activities require subject-specific training for cross-day and cross-environment recognition.Standing-up and sitting-down movements produce nearly indistinguishable Doppler traces; separating the new activities requires finer, subject-specific traits.Source: §6.7 · pp. 13–14, Table 6 · p. 14

• The system analyzes Wi-Fi signals on commonly available COTS routers, avoiding a separate sensing infrastructure for deployment and maintenance.Reliable channel frequency response estimation is possible while receiving background data traffic.Source: §7 · p. 14
• SHARP quantifies Doppler shifts from monitored CSI and classifies activity-specific traces, which are not affected by environment geometry or static objects.Once trained, it can be used at runtime on unseen environments and individuals without complex, periodic calibration.Source: §7 · p. 14
• 96% average accuracy in the most challenging scenario, where both person and environment differ from training; SHARP also outperforms state-of-the-art CSI-amplitude- and phase-based methods.Source: §7 · p. 14
• Above 95% accuracy was maintained for additional similar-dynamics activities in an environment different from training.These activities include standing, sitting down or standing up, and arm gymnastics.Source: §7 · p. 14
• Small errors occur with unknown persons and when distinguishing walking from running, which the authors attribute to behavioral differences or similar movements.Source: §7 · p. 14
• Future work includes cross-hardware operation and recognizing multiple people moving concurrently.The authors note practical implications for real deployments and applications.Source: §7 · p. 14

• The appendix details the main blocks of a Wi-Fi transmission chain to derive the channel frequency response in Eq. (2).Source: Appendix A · p. 15

• OFDM transmits closely spaced orthogonal subchannels in parallel, each modulated with a conventional digital modulation scheme.Source: §A.1 · p. 15
• Input bits map to complex constellation symbols, which are grouped into OFDM symbols and transformed by an IFFT for transmission over K subchannels.The subchannel spacing is ∆f = 1/T Hz, where T is the OFDM symbol time.Source: §A.1 · p. 15
• For IEEE 802.11ac, the 80 MHz transmission uses K = 256 subchannels, ∆f = 312.5 kHz, T = 3.2 µs, and a 4 µs OFDM symbol including cyclic prefix.The cyclic prefix duration is TCP = 0.8 µs and is added to mitigate inter-symbol interference.Source: §A.1 · p. 15
• The m-th OFDM symbol is represented by the vector of its K subchannel samples, and the model considers M subsequent blocks before upconverting x(t) to carrier frequency fc.Source: §A.1 · p. 15

• Each receiver antenna collects P signal copies from multipath scatterers, with each path characterized by attenuation Ap(t) and delay τp(t).Source: §A.2 · p. 15
• A rectangular receiver window collects one OFDM symbol, and a Fourier transform of the received window recovers transmitted symbol ak.The model assumes path attenuation and delay are constant over an integration interval T.Source: §A.2 · p. 15
• The sum in the frequency-domain expression corresponds to the Wi-Fi channel frequency response, estimated from known preamble symbols.Source: §A.2 · p. 15

A.3 Phase offsets in the Wi-Fi channel estimates

• Hardware artifacts cause CFR measurements from Wi-Fi devices to deviate slightly from the model by introducing phase offsets, or rotation errors.The passage identifies these as the most significant artifacts affecting phase information.Source: §A.3 · p. 15
• CFO, SFO, and PDD produce offsets from carrier-frequency mismatch, transmitter–receiver clock desynchronization, and packet-detection delay, respectively.CFO is only partially compensated for at the receiver.Source: §A.3 · p. 15
• PPO arises from the PLL's random generation of the transmitter's initial phase, while phase ambiguity reflects antenna phase differences in multiples of π.In static conditions, the phase difference between antennas should remain constant.Source: §A.3 · p. 16
• CFO, SFO, and PDD contributions are shared across antennas, whereas PPO and phase ambiguity are antenna-specific.The passage attributes this distinction to the antenna-specific initial PLL phase and phase ambiguity.Source: §A.3 · p. 16