Abstract
Indoor navigation (IN) systems are increasingly essential in environments where GPS signals are unreliable, such as hospitals, airports, and large public buildings. This study explores a smartphone-based approach to indoor positioning that leverages inertial sensor data for accurate step detection and counting, which are fundamental components of pedestrian dead reckoning. A long short-term memory (LSTM) network was trained to recognize step patterns across a variety of indoor movement scenarios. The generalized model achieved an average step detection accuracy of 93%, while scenario-specific models tailored to particular movement types such as turning, stair use, or interrupted walking achieved up to 96% accuracy. The results demonstrate that incorporating activity-specific training improves performance, particularly under complex motion conditions. Challenges such as false positives from abrupt stops and non-walking activities were reduced through model specialization. Although the system performed well offline, real-time deployment on mobile devices requires further optimization to address latency constraints. The proposed approach contributes to the development of accessible and cost-effective indoor navigation systems using widely available smartphone hardware and offers a foundation for future improvements in real-time pedestrian tracking and localization.