Real-time energy-efficient indoor tracking system based on edge AI and ultra-low-power sensor nodes
Mohamed Khalil Baazaoui, Robert Fromm, Ahmed Fakhfakh, Faouzi Derbel · Results in Engineering · 2026
Indoor localization is critical for applications such as industrial automation, asset tracking, and human navigation in hospitals, airports, and shopping centers. However, existing approaches based on WiFi fingerprinting, wireless sensor networks, and MEMS-based inertial systems suffer from infrastructure dependency, network congestion, high energy consumption, and cumulative drift that degrades long-term accuracy. This work presents an ultra-low-power hybrid indoor localization architecture for real-time on-device position estimation on a resource-constrained System on Chip (SoC). The proposed system combines inertial sensing with event-driven anchor-based ranging enabled by a Wake-Up Receiver (WuRx). The WuRx consumes only 15.7 μ W, allowing the sensor node to remain in an ultra-low-power sleep state until activation, while the tracking node achieves 35.7 μ W sleep-mode power consumption. A tightly coupled fusion framework integrates adaptive Zero-Velocity Updates (ZUPT) and Zero-Angular-Rate Updates (ZARU), robust heading correction, and probabilistic scale estimation to mitigate drift and improve long-term stability. Unlike conventional approaches relying on continuous ranging or centralized processing, the proposed architecture performs the complete processing pipeline on-device using an edge-computing paradigm. By combining inertial tracking with event-driven ranging updates, it reduces communication overhead and minimizes the number of required active localization nodes while maintaining reliable position estimation. To compensate for inertial bias and scale drift, both Weighted Least Squares (WLS) and Weibull-based Bayesian correction methods were investigated. The Weibull-Bayesian approach achieved superior performance with accurate trilateration error modeling, yielding localization Root Mean Squared Error (RMSE) of 1.5 m over a 205.17 m indoor trajectory while consuming 0.0048% of battery capacity. Trilateration and fusion require only 0.9 ms per update, enabling real-time operation.