TLDR: This research introduces a robust and accurate method for tracking daily paths of inhabitants in home environments using Ultra-Wideband (UWB) technology and deep learning models. By employing an RSSI fingerprinting approach, the system overcomes limitations of traditional UWB and Bluetooth Low Energy (BLE) methods, which are often affected by obstacles. The study demonstrates that a hybrid CNN+LSTM deep learning model, processing UWB RSSI data, can achieve a mean absolute error close to 50 cm in real-world home settings, providing superior location estimates and facilitating human activity recognition while being cost-effective and privacy-preserving.
Understanding where people are and what they are doing inside their homes is becoming increasingly important for smart home systems, healthcare, and assisted living. This field, known as human activity recognition (HAR), has seen significant advancements thanks to new Internet of Things (IoT) devices. However, accurately tracking individuals indoors, especially in multi-resident homes, remains a challenge.
Traditional methods for indoor positioning often rely on technologies like Bluetooth Low Energy (BLE) or Ultra-Wideband (UWB). While BLE is cost-effective and widely available in devices, its signal strength can be inconsistent due to furniture and other obstacles, requiring many beacons for accuracy. UWB offers higher precision, but its signals are also affected by walls and obstructions, and it typically needs many strategically placed anchors with a clear line of sight, which can be expensive and complex to deploy.
A recent study proposes a novel approach to overcome these limitations by using UWB technology combined with a technique called ‘fingerprinting’ and advanced deep learning models. Instead of relying on traditional methods that calculate location based on signal travel time (which struggles with obstacles), this research uses Received Signal Strength Indicator (RSSI) data. RSSI measures the power of a received radio signal, and by mapping these signal strengths at various locations, a ‘fingerprint’ of the environment is created.
The Proposed System
The researchers developed a system that deploys UWB anchors on walls and ceilings within a home. Users wear small UWB tags that send RSSI data to these anchors. This data is then sent to a server where deep learning models process it. The system learns from a dataset where users manually label their real-time locations on a map, correlating these labels with the RSSI signals received by the anchors. Once trained, the system can accurately estimate a person’s location in real-time using only the RSSI signals.
The study compared the performance of different deep learning models, including Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM) networks, and a hybrid CNN+LSTM model. CNNs are good at recognizing spatial patterns in data, while LSTMs excel at understanding temporal dependencies in signal sequences. The hybrid model combines these strengths to provide more sophisticated and adaptive location estimates.
Real-World Testing and Results
To test their approach, the researchers conducted case studies in two different flats, one 60 square meters and another 100 square meters. They collected extensive data from inhabitants performing daily activities, with locations meticulously labeled in real-time using a tablet application. They also compared UWB performance with BLE in these real-world conditions.
The results were promising. For room classification, UWB RSSI data achieved 85% accuracy in Flat A and 83% in Flat B, significantly outperforming BLE (70% and 75% respectively). For precise location estimation (X and Y coordinates), the hybrid CNN+LSTM model consistently delivered the best performance. The mean absolute error (MAE) for location estimation was close to 50 cm, demonstrating high accuracy. The study also found that increasing the ‘window size’ (the duration of temporal data considered by the models) generally reduced the error, though with diminishing returns.
Crucially, the fingerprinting-based approach using RSSI and deep learning proved more robust than traditional UWB trilateration, especially in areas with poor signal coverage where trilateration often failed to provide a location. This highlights the system’s ability to work effectively even in complex indoor environments with obstacles.
Also Read:
- UrbanMIMOMap: A New Dataset for 6G Radio Map Research
- Deep Learning Powers Next-Generation Integrated Sensing and Communication
Future Directions
This research demonstrates the potential of combining UWB RSSI fingerprinting with deep learning for accurate and cost-effective indoor location tracking. The system offers a low computational demand and respects privacy, as it doesn’t rely on cameras or intrusive sensors. The authors plan to deploy this architecture in a 24/7 operational environment to gather more data and further refine their models. They also aim to integrate ambient binary sensors to enhance human activity recognition in multi-occupancy settings and explore edge computing approaches where UWB scanning is integrated directly with wearable devices, potentially reducing infrastructure needs. For more details, you can read the full research paper here.
