TL;DR
ImmTrack is a privacy-preserving system that fuses millimeter-wave (mmWave) radar with smartphone IMU data to track interpersonal distances with decimeter-level accuracy. By matching movement "trace maps" between the two sensors, it solves the radar's anonymity problem, enabling effective contact tracing that outperforms Bluetooth and rivals camera-based systems without the privacy risks.
What Problem Does This Paper Address?
Effective contact tracing for infectious diseases like COVID-19 requires "decimeters-seconds" resolution—knowing exactly how close people were and for how long. Current solutions fall into two camps, both flawed:
- Bluetooth Neighbor Discovery (BND): The most common approach, but it suffers from high latency and inaccurate distance estimation due to signal multipath and attenuation. It often fails to detect contacts shorter than five minutes.
- Device-Free Sensing (Cameras/Radar): Cameras are highly accurate but raise significant privacy concerns. mmWave radar is a great middle ground—it provides high-resolution 3D data without identifying faces—but it suffers from an anonymity problem. It can track "blobs" moving through a room, but it doesn't know which blob corresponds to which person (identity).
The authors aim to bridge this gap: providing the high-precision tracking of radar while using the "pseudo-identities" inherent in a user's smartphone data to label those tracks.
Methods
The authors introduce ImmTrack, which employs a two-stage fusion process:
- IMU-Assisted mmWave Tracking: To prevent the radar from "merging" two people into one cluster when they walk close together, ImmTrack uses Recursive Kalman Filters (RKF) to predict where each person will be in the next frame. This "motion-aware" clustering initializes a $k$-means algorithm, ensuring clusters remain distinct even during near-contacts.
- mmClusterNet: A custom deep neural network extracts shape and motion features from these radar point clouds to maintain consistent tracking across frames.
- Trace Maps & Siamese Networks: This is the core innovation for cross-modality matching. The system converts 2D trajectories from both the radar and the smartphone's IMU (via dead reckoning) into "trace maps"—images where pixel intensity represents the time a person spent at a location. A Siamese Neural Network then compares these maps to find the best match between a radar track and a specific smartphone ID.
- Gait Analysis Fallback: If two people walk side-by-side (creating nearly identical trace maps), the system falls back to gait analysis, comparing the Doppler spectrogram from the radar against the acceleration patterns from the IMU to tell them apart.
Datasets
The researchers evaluated ImmTrack through several real-world deployments:
- Controlled Environments: Experiments in a sports hall and an outdoor playground with up to 7 people to establish baseline accuracy.
- Live Lab Space: A four-day deployment in a 300 $m^2$ office/lab environment with 17 residents and various "passengers" (untracked visitors), totaling up to 27 people.
- Training Data:
mmClusterNetwas trained using the public ShapeNet dataset and self-collected radar data. The Siamese network was trained using a mix of real traces and simulated trajectories generated via random walk stochastic processes for data augmentation.
Key Insights and Ideas
- Identity via Movement: The fundamental "aha" moment is that while radar data is anonymous, a person's movement path is a unique fingerprint over short windows of time. By matching the "how" (movement) between the radar and the phone, you can transfer the "who" (ID) to the "where" (radar coordinates).
- Modality-Agnostic Representation: Converting trajectories into "trace maps" (images) allows the authors to treat a complex sensor-fusion problem as a computer vision problem. This bypasses the need for perfect coordinate system registration between the radar and the phone's internal sensors.
- Motion-Awareness is Mandatory: Standard clustering (like DBSCAN) fails when people touch or cross paths. By making the clustering step "aware" of the previous frame's velocity and position, the system handles the very "close contacts" it is designed to monitor.
Key Results
- Spatial Accuracy: ImmTrack achieved an average interpersonal distance tracking error of just 22 cm.
- Contact Detection: For pinpointing contacts within 1 meter, the system achieved 90% precision and 94% recall.
- Latency: Compared to Bluetooth-based systems, ImmTrack reduced discovery latency by up to 80 seconds, making it viable for detecting "fleeting encounters."
- Efficiency: The system provides tracking accuracy similar to camera-based systems (ICTrack) but requires only 1/4 to 1/2 of the computational power, making it more suitable for edge computing deployments.
Limitations and Future Work
- Enrolment Requirement: Users must still have an app installed and "sign in" to the space (e.g., via QR code), which may limit adoption in purely public spaces.
- Hardware Deployment: Unlike Bluetooth, this requires the installation of specialized mmWave radar hardware in the environment.
- Occlusion: While mmWave radar handles some occlusion better than cameras, dense crowds or complex architectural "blind spots" still pose challenges that require careful multi-radar placement.
- Future Directions: The authors suggest exploring how to handle even more dynamic environments and potentially integrating other low-power sensors to further reduce the reliance on active smartphone transmission.
Summary updated