ELF-SLAM: Indoor Smartphone SLAM with Learned Echoic Location Features

Opportunity

Indoor self-localisation is a highly demanded function for smartphones. As of 2022, over 70% of the top 100 free Android apps require location information. Current solutions relying on Wi-Fi, GSM, FM radio and visible light have limitations such as susceptibility to electromagnetic noise, signal blockage and privacy concerns with visual imaging. ELF-SLAM addresses these issues by using learned echoic location features from smartphone-collected echoes. Echoes carry sub-metre granularity location information, enabling more accurate sensing than traditional methods. As an infrastructure-free system, ELF-SLAM does not require anchor points and exhibits great potential for commercial production.

Technology

ELF-SLAM operates in two phases: mapping and localisation. Figure 1 gives an overview of the mapping phase. During mapping, as the user moves with a smartphone, both inertial measurement unit (IMU) data and acoustic echoes are collected. Loop closures, detected using echoes, correct IMU-based dead reckoning deviations, resulting in accurate trajectory maps. In the localisation phase, the smartphone’s location is determined using captured echoes and the pre-constructed map.

ELF-SLAM includes three key designs. The trajectory-level contrastive learning (CL) procedure learns ELF for loop closure detection. Second, a clustering-based approach curates loop closures, effectively removing false positives. Third, a floor-level CL procedure superimposes the trajectory maps from multiple users into a single floor map, reconciling differences in ELFs due to smartphone orientation.

Figure 1. Overview of the mapping phase of ELF-SLAM.

Applications & Advantages

Main application areas include indoor navigation for smartphones, enhancing location-based services in homes, offices and large commercial spaces.

Advantages:

• Does not require additional hardware installation.
• Achieves sub-metre location sensing in various indoor spaces.
• Employs contrastive learning to derive echoic location features from a few minutes of unlabelled echo data.
• Proven accuracy in diverse environments, including small homes (~60m²), medium offices (~350m²) and large shopping malls (~1200m²).