Sensor-Fusion for Smartphone Location Tracking Using Hybrid Multimodal Deep Neural Networks

Area of Science:

• Computer Science
• Machine Learning
• Signal Processing

Background:

• Indoor localization using smartphone sensors is a complex challenge.
• Existing engineered solutions struggle with edge cases.
• Data-driven approaches offer potential but require specialized architectures.

Purpose of the Study:

• To propose an end-to-end hybrid multimodal deep neural network for indoor localization.
• To develop a system that learns features automatically from data, eliminating the need for hand-engineering.
• To address the limitations of traditional indoor positioning systems in difficult environments.

Main Methods:

• Developed MM-Loc, a hybrid multimodal deep neural network system.
• Utilized modality-specific neural networks for feature extraction from diverse sensors (inertial, magnetic, WiFi).
• Employed cross-modality neural structures to fuse extracted features for enhanced localization accuracy.

Main Results:

• MM-Loc demonstrated superior performance compared to traditional localization approaches.
• The system effectively handles cross-modality data with varying sampling rates and representations.
• Independent modality-specific networks provided location estimates, but multimodal fusion yielded better accuracy.

Conclusions:

• MM-Loc offers an effective, data-driven solution for indoor localization challenges.
• The hybrid multimodal deep learning approach surpasses conventional methods, especially in complex scenarios.
• Automatic feature learning from sensor data represents a significant advancement over human-engineered systems.