CNN-Based Real-Time Walking Direction Estimation for Pedestrian Navigation Scenarios

In pedestrian navigation scenarios, the walking direction of pedestrians is usually assumed to be the same as the upward direction of the screen of their handheld mobile devices. However, the two aforementioned directions are not always guaranteed to be the same, as pedestrians hold their devices arbitrarily. This study investigates a deep learning-based approach that estimates the difference between the two directions, called the heading offset, to improve the reliability of several pedestrian applications, including navigation. To this end, we propose a convolutional neural network (CNN) model that considers accelerometer and GYRO measurements and produces an estimated heading offset value. In order to reduce the labor required for collecting training data, we use an automatic annotation technique that acquires the ground truth labels when global positioning system (GPS) coordinates are reliable, and we also propose a data augmentation method. The performance of the proposed method was extensively evaluated with data collected in both indoor and outdoor environments, and it was confirmed that the proposed CNN-based approach outperforms existing model-based approaches. In addition, we implemented an application to demonstrate that the proposed method is executed in commercial mobile devices in real-time.