Machine Learning-Driven Indoor Positioning: A New Paradigm of Intelligent Localization Beyond Traditional Methods

This article introduces an undergraduate thesis project in computer science that explores using machine learning models to go beyond traditional indoor positioning methods, compensating for the impacts of network constraints and environmental changes by learning environmental patterns. The project's core adopts a four-stage pipeline architecture, combining traditional positioning methods with machine learning technologies, aiming to provide a new path for device positioning in indoor-outdoor hybrid environments, with wide practical application value.

This project is sponsored by Clear-Com Telecommunications, a company that provides hardware and embedded software solutions for professional fields such as entertainment, nuclear facilities, military, and space exploration. Since its devices are often in indoor-outdoor hybrid environments, accurate device positioning is crucial to its business. This problem is universal—many organizations face similar challenges—making the project's results widely applicable.

The project adopts a four-stage closed-loop architecture:

1. Network Environment Simulation and Data Generation: Simulate various 2D network environments, generate diverse data including location, obstacles, LOS/NLOS status, etc., to ensure the diversity and relevance of the dataset.
2. Multi-Method Position Estimation: Use three mainstream methods—RSSI (signal strength), TDOA (time difference), DOA/AOA (angle)—to calculate target positions, set grid constraints, and simulate real signal conditions.
3. Machine Learning Model Training: Map positioning results with environmental data to form a training set, using incremental training and regularization techniques to prevent overfitting.
4. Model Evaluation and Optimization: Compare performance with traditional methods; if no significant improvement is achieved, adjust its role to assist traditional methods.

The project has sufficient technical challenges: it involves the mathematical complexity of positioning methods, network behavior modeling and simulation, multi-method evaluation, and model design, reflecting the ability to solve problems independently and conduct critical analysis. The literature foundation includes: Alawieh & Kontes (5G positioning combined with AI/ML), Rathnayake et al. (RSSI + ML for smart city positioning), Rajput et al. (ML sample size evaluation), and Xie et al. (multipath suppression for WiFi positioning).

The core innovation lies in breaking through the limitation of traditional methods that treat environmental factors as noise, instead taking environmental features as learnable signals. Practical significance:

• Provide a new path for positioning in indoor-outdoor hybrid environments
• Enhance the robustness of positioning systems against environmental changes
• The method is universal and scalable to multiple scenarios and industries

This project represents an important attempt in the evolution of indoor positioning toward intelligence and adaptability. By combining traditional methods with machine learning, it not only improves positioning accuracy but also provides new ideas for dealing with complex environments. With the popularization of 5G and the surge in IoT devices, high-precision indoor positioning will play a key role in fields such as smart buildings, industrial automation, and emergency response.