A Multiscale Approach to Detect the Semantics of Locomotion without Positioning Information

The detection of locomotion patterns and trajectory semantics has long relied on the positioning information. However, positioning information can be unavailable in specific scenarios. In addition, the position values are sensitive to rotation and hence a bias can exist during the training process in deep learning models. This paper introduces an alternative model that classifies trajectories based on accelerometers, instead of positioning systems. We built up a convolutional neural network that inputs the degree of velocity and turning angles at several time scales and converts this information into a semantic class. We examined the model in a simulated environment and also a benchmark task. This model has exhibited a competitive performance even compared with those models based on positioning information.