Enhancing GPS Data Accuracy in SAP Systems Using IMU Sensors and Machine Learning

The Strava platform received 6.6 billion miles' worth of activity data in 2018. If consumers want to get fitter faster than their peers or professional athletes—they need accurate data. However, because Global Positioning System (GPS) data is noisy and low-frequency—users are sometimes frustrated by false paths—that suggest—they have ridden across lakes or sprinted through buildings. Therefore, this study uses sensors, such as accelerometers, gyroscopes, and magnetometers—found in an Inertial Measurement Unit (IMU) to address this issue. We look into how to counteract the drift of high-frequency IMU sensors over time and low-frequency GPS sensors. Real-time correction during the activity—which necessitates computational efficiency, and post-activity correction—which permits the utilization of previous and upcoming measurements—are the two use cases that are examined. For these use cases, two approaches are used— Recurrent Neural Networks (RNNs), a recent technology, and the Kalman Filter, a traditional technique from the 1960s. The Kalman Filter can do a reverse pass post-activity for more accurate estimations by combining sensor values with historical positions. Without requiring specific sensor variance information, RNNs—such as Gated Recurrent Units and Long Short Term Memory Networks—use built-in memory modules to find patterns in sensor data. RNNs outperformed the Kalman filter in producing qualitatively credible findings, according to both approaches. However, inconsistent ground truth data made quantitative evaluation difficult. Prospective investigations are suggested to address these constraints.