Abstract: Radio Frequency Identification (RFID) technology has been widely used in many application domains. How to apply RFID technology to develop an Indoor Positioning System (IPS) has become a hot research topic in recent years. LANDMARC approach is one of the first IPSs by using active RFID tags and readers to provide location based service in indoor environment. However, major drawbacks of the LANDMARC approach are that its localization accuracy largely depends on the density of reference tags and the high cost of RFID readers. In order to overcome these drawbacks, two localization algorithms, namely weighted path loss (WPL) and extreme learning machine (ELM), are proposed in this paper. These two approaches are tested on a novel cost-efficient active RFID IPS. Based on our experimental results, both WPL and ELM can provide higher localization accuracy and robustness than existing ones.
Introduction
While GPS has mastered outdoor navigation, indoor positioning remains a significant challenge. Satellite signals cannot penetrate walls, and indoor environments are chaotic—filled with furniture that scatters signals, moving people, and "multipath effects" where signals bounce off surfaces.
Radio Frequency Identification (RFID) is a leading candidate for indoor tracking because tags are inexpensive and do not require a line-of-sight. However, the industry-standard LANDMARC approach has two major flaws: it requires expensive RFID readers and its accuracy depends heavily on a high density of "reference tags." This paper presents a significant step forward by moving away from expensive proprietary hardware toward clever machine learning and signal modeling, demonstrating a system that outperforms industry benchmarks.
Methodology and Experimental Setup
The researchers developed a cost-effective RFID indoor positioning system using sensors based on the TICC2530 wireless module, costing approximately $15 each. The system was tested in a 6.4m × 17.1m "Postgraduate Room" at Nanyang Technological University.
The setup utilized 19 RFID sensors, 19 reference tags, and 7 tracking tags. Data collection spanned 19 days, generating over 1.2 million RSSI samples to test two distinct algorithmic approaches:
- Weighted Path Loss (WPL): A centralized, model-based approach. It uses a modified International Telecommunication Union (ITU) indoor path loss model to convert Received Signal Strength Indicator (RSSI) values into distances. It then calculates a "weighting factor" for each sensor to estimate coordinates.
- Extreme Learning Machine (ELM): A "fingerprinting" approach based on a Single-hidden Layer Feedforward Neural Network (SLFN). Unlike traditional neural networks, ELM randomly assigns input weights and only "learns" output weights using a simple least-squares method. This makes it incredibly fast to train while remaining highly effective at mapping complex, non-linear signal patterns.
Key Results and Discussion
The researchers compared their methods against LANDMARC and an "Enhanced LANDMARC" variant. The results showed that the ELM-based approach significantly improved localization accuracy.
- Accuracy: LANDMARC resulted in a 2.642m average error, while WPL improved this to 1.651m. The ELM approach (with 2000 nodes) achieved the best performance with a 1.198m average error—a 55% improvement over LANDMARC.
- Robustness: One of the most impressive insights is that ELM maintains high accuracy even when nearly half of the signal packets are lost or corrupted. In tests with "defect" samples, ELM maintained an error of 1.137m, while LANDMARC’s error ballooned to 3.503m.
- Efficiency: A clear trade-off exists between accuracy and speed. While ELM was the most accurate, WPL was significantly faster, processing samples in under 0.2 seconds, whereas the high-accuracy ELM (2500 nodes) took nearly 2 seconds.
Conclusion
This research proves that high-end readers are not strictly necessary for accurate indoor positioning. By using $15 ZigBee-based sensors and sophisticated algorithms like ELM and WPL, researchers can achieve better results than systems using hardware costing ten times as much. Future work will focus on hybridizing the two approaches—potentially using the speed of WPL for initial estimates and the precision of ELM for fine-tuning—to create a system that is both lightning-fast and highly accurate.
Summary updated