2D Orientation Estimation Using Machine Learning With Multiple 5G Base Stations

Abstract: Localization of mobile devices has implications on a multitude of use cases such as estimating the location of the user originating an emergency call, localization of devices to enable autonomous operation required by industrial Internet of Things (IoT) use cases, etc. In futuristic use cases such as Augmented Reality (AR), Virtual Reality (VR), Extended Reality (XR), autonomous navigation of Unmanned Aerial Vehicles (UAVs), we will require the capability of estimating orientation in addition to position of such devices for efficient and effective provisioning of these services to the end-users. One way to handle the problem of finding the orientation of devices is to rely on the measurements from different sensors like the magnetometer, accelerometer and gyroscope but the limitation of this method is the dependency on these sensors, and thus cannot be used for some devices which does not have these sensors. Hence these limitations can be overcome by using data-driven approaches like Machine Learning (ML) algorithms on received signal features, where a training dataset with orientation measurements are used to train the ML model that can transform the received signal measurements to orientation estimates. The data for the work is generated by using simulator that can simulate the environment with multiple base stations and receivers. The measurements or features that are generated from the simulator are the Received Signal Received Power (RSRP), Time of Arrival (ToA), Line of Sight (LoS) condition, etc. In-order to find the relationship between the received signal features and orientation, two nonlinear ML algorithms namely K Nearest Neighbors (KNN) and Random Forest (RF) are used. The received measurements were investigated and RSRP was identified as the feature for the ML models. The ML algorithms are able to estimate the orientation of the User Equipment (UE) by using KNN and RF, where different features likes RSRP and the information about LoS and Non Line of Sight (NLoS). These features were used alone and also combined to evaluate the performance. The results also shows how interference of radio signals affects the performance of the model. Adding to that, different combination of received signal features were also used to compare the performance of the model. Further tests were also done on the trained model to identify how well it can estimate orientation when a new UE with new position is introduced.