SPSTS-Pos: AI-driven indoor positioning with strongest-path space-time structural features in O-RAN system

With the adoption of larger bandwidths and massive antennas, mobile cellular networks exhibit substantial potential for supporting positioning functionalities by leveraging shared hardware resources. Based on this architecture, this paper presents a high-precision indoor positioning method for multiantenna 5G systems, termed as SPSTS (Strongest-Path SpaceTime Structure). Unlike traditional geometric or fingerprintbased positioning approaches, SPSTS extracts relative timedelay differences of the strongest multipath component (MPC) across antenna pairs using uplink SRS signals. Based on the extracted SPSTS data, we have designed a customized ResNet architecture with asymmetric convolutions to decouple spatiotemporal features and regress them into 2D coordinates. The method is implemented on the O-RAN-compliant platform using an OpenAirInterface BBU, a commercial O-RU and distributed antennas, operating 100 MHz bandwidth in an indoor office environment. Real-world experiments achieve state-of-the-art positioning accuracy of 0.21 m (CDF50%) in offline evaluation and 0.35 m (CDF90%) in real-time testing. A video demonstrating the real-time positioning is available at www.opensource5g.org/demo/SPSTS-Positioning.mp4.