Abstract
Accurate indoor positioning for unmanned aerial vehicles (UAVs) is critical for logistics, surveillance, and emergency response applications, particularly in GPS-denied environments. Existing indoor localization methods, including optical tracking, ultra-wideband, and Bluetooth-based systems, face cost, accuracy, and robustness trade-offs, limiting their practicality for UAV navigation. This paper proposes CiUAV, a novel 3D indoor localization system designed for UAVs, leveraging channel state information (CSI) obtained from low-cost ESP32 IoT-based sensors. The system incorporates a dynamic automatic gain control (AGC) compensation algorithm to mitigate noise and stabilize CSI signals, significantly enhancing the robustness of the measurement. Additionally, a multi-task 3D localization model, Sensor-in-Sample (SiS), is introduced to enhance system robustness by addressing challenges related to incomplete sensor data and limited training samples. SiS achieves this by joint training with varying sensor configurations and sample sizes, ensuring reliable performance even in resource-constrained scenarios. Experiment results demonstrate that CiUAV achieves a LMSE localization error of 0.2629 m in a 3D space, achieving good accuracy and robustness. The proposed system provides a cost-effective and scalable solution, demonstrating its usefulness for UAV applications in resource-constrained indoor environments.
Introduction
Unmanned Aerial Vehicles (UAVs) are increasingly deployed in complex indoor environments such as warehouses, factories, and disaster zones. In these settings, the absence of GPS signals necessitates alternative localization technologies. While optical tracking and Ultra-Wideband (UWB) systems offer high precision, they are often hindered by high costs, strict line-of-sight requirements, or the need for specialized hardware on the UAV itself.
WiFi Channel State Information (CSI) has emerged as a promising alternative, providing fine-grained signal characteristics that can be used for device-free or passive tracking. However, CSI-based 3D localization faces significant hurdles: signal noise caused by hardware-level Automatic Gain Control (AGC), the high dimensionality of 3D space, and the requirement for massive labeled datasets to train robust deep learning models. CiUAV addresses these challenges through a hardware-software co-design approach.
System Architecture
The CiUAV system is designed to track drones without requiring any additional hardware on the aircraft, preserving battery life and payload capacity. The architecture consists of three primary stages: data acquisition, signal compensation, and multi-task localization.
Passive CSI Sniffing
The system utilizes low-cost ESP32-S3 IoT sensors acting as passive sniffers. These sensors send standard ICMP Echo Requests (Pings) to the drone. When the drone responds with standard WiFi packets, the sensors extract the CSI from the physical layer. This approach ensures compatibility with any WiFi-enabled UAV.
Dynamic AGC Compensation (DAC)
A major source of error in CSI amplitude data is the receiver's Automatic Gain Control (AGC), which dynamically adjusts signal amplification to maintain connection stability. This hardware behavior distorts the relationship between signal strength and distance. CiUAV extracts real-time AGC gain values directly from the ESP32 firmware and applies a linear scaling factor to mathematically reverse this distortion, restoring the true signal amplitude before it reaches the neural network.
Sensor-in-Sample (SiS) Multi-Task Model
To predict 3D coordinates $(x, y, z)$, the system employs a Multi-Task Localization Network (MTLN). Traditional models often fail if a single sensor drops offline. The SiS approach jointly trains the network across multiple "tasks," such as localization using different subsets of sensors and varying sample sizes. By sharing a feature extraction backbone (e.g., MobileNet or ResNet), the model learns generalized spatial features that are resilient to sensor dropouts and data scarcity.
Experimental Evaluation
The system was validated in a 5m × 5m carpeted indoor laboratory. Three ESP32 CSI sensors were ceiling-mounted to provide spatial diversity.
Data Collection and Ground Truth
The UAV was flown to 121 predefined 3D grid points at heights ranging from 0.6m to 2.0m. Ground truth positioning was established using an onboard camera that recognized Aruco markers on the floor, combined with a Time-of-Flight (ToF) sensor for precise altitude detection. The resulting dataset included 110,000 samples (77,000 for training and 33,000 for testing) at a 50Hz sampling rate.
Results and Key Insights
CiUAV demonstrates state-of-the-art performance for low-cost CSI-based 3D tracking:
- High Precision: Using a MobileNet-based MTLN, the system achieved a Localization Mean Squared Error (LMSE) of 0.2629 meters.
- Impact of DAC: Ablation studies showed that the Dynamic AGC Compensation module reduced the LMSE by approximately 5.7 cm, highlighting the importance of hardware-aware signal processing.
- Data Efficiency: The SiS model proved remarkably efficient; even when trained on only 25% of the available data, the localization error only increased slightly from 0.51m to 0.56m.
- Robustness: The multi-task training allowed the system to maintain functional tracking (1.33m error) even when two out of three sensors were simulated as offline, a scenario where single-task models typically fail to produce valid outputs.
Conclusion
CiUAV provides a robust, scalable, and cost-effective solution for 3D indoor UAV localization. By combining hardware-level AGC compensation with a resilient multi-task learning framework, the system overcomes the traditional limitations of CSI-based tracking. Future research will focus on adapting the system to larger, more dynamic environments with moving obstacles and integrating temporal models to track continuous flight trajectories.
Summary updated