Abstract
DynaPose is a deep-learning-based system designed to enable seamless "tagless" entry gates, such as subway turnstiles, that operate without requiring users to remove their smartphones from their pockets. By integrating Ultra-wideband (UWB) signal diagnostics with smartphone motion sensors, DynaPose identifies signal blockages caused by the human body and dynamically selects optimal anchors. This approach improves localization accuracy by 62% in challenging Non-Line-of-Sight (NLOS) conditions, ensuring reliable and timely gate operations.
Introduction
The vision of proximity-based services involves doors and fare gates unlocking automatically as a user approaches. Ultra-wideband (UWB) technology is a primary candidate for this due to its centimeter-level precision. However, real-world deployment is hindered by the human body, which effectively absorbs UWB signals. When a device is in a pocket, the body creates NLOS conditions, leading to signal attenuation and multipath errors. These issues cause gates to malfunction—opening too early, too late, or not at all.
System Architecture
DynaPose addresses these challenges through a multi-stage real-time framework running on a smartphone. The system transitions from wide-area localization to precise gate-access ranging.
1. Effective CIR (eCIR) Updater
Standard UWB Channel Impulse Response (CIR) data is too large for rapid mobile transfer. DynaPose uses an "eCIR" updater that crops data to 135 relevant samples around the "First Path Index," reducing latency from 223.4ms to 17.8ms.
2. LOS/NLOS Classification
A 1D Convolutional Neural Network (CNN) analyzes the eCIR to determine if the signal path is clear (Line-of-Sight, LOS) or obstructed (NLOS).
3. Dynamic Anchor Selection Algorithm (ASA)
Upon detecting NLOS, the system employs a Graham Scan algorithm to calculate a convex hull of available anchors. It selects anchor combinations that maintain high signal quality while ensuring the user remains within the geometric "fold" of the anchors for optimal trilateration.
4. Pose Prediction
A hybrid CNN-LSTM architecture combines UWB signal signatures with Inertial Measurement Unit (IMU) data. This allows DynaPose to distinguish between four poses: Hand-LOS, Hand-NLOS, Front-Pocket, and Back-Pocket, by capturing both radio multipath signatures and physical motion patterns.
Experimental Evaluation
The system was evaluated using a Samsung Galaxy Note20 Ultra and a Qorvo MDEK1001 UWB module. The dataset included 50,000 UWB samples (8,000 custom, 42,000 open-source) and 26,400 IMU samples across various poses and indoor environments.
Key Results
- Accuracy: The system achieved 98.4% accuracy in LOS/NLOS classification and 96.1% in real-time pose detection.
- Localization Performance: In "back pocket" NLOS scenarios, dynamic anchor selection improved localization accuracy by 62% over legacy systems.
- Latency: All processing and transitions occurred within one second, meeting the requirements for pedestrian traffic flow.
Conclusion and Future Work
DynaPose demonstrates that combining radio diagnostics with motion sensing can overcome the physical limitations of UWB in mobile environments. Current limitations include the need for external hardware due to restricted smartphone UWB APIs. Future research will focus on high-density multi-user environments and optimizing power consumption for continuous mobile operation.
Summary updated