IMU Based Context Detection of Changes in the Terrain Topography

Abstract

This paper introduces an IMU based context machine learning algorithm for terrain topography classification. Four different terrains are considered: concrete, pebble, sand, and grass. The grass terrain is further split into two separate classes based off moisture content of the grass, wet and dry. Separate terrain topography datasets are created by walking on different terrains and logging the data. The subject has been equipped with an IMU attached on the surface of the shoe above the toes. Data is collected and stored via a Bluetooth smartphone controller over multiple recording sessions. Acceleration, angular rate, and magnetic field were recorded. The recorded data is extracted in two second sliding window intervals, whereupon the magnitude of the sensor outputs, in three dimensions, is calculated. A low-pass band filter is also applied to the magnitude for the acceleration, angular rate, and magnetic field data. The magnitude output is processed in the time domain to calculate variance, energy, kurtosis, range, skewness, and the zero-crossing rate. The magnitude data is converted into the frequency domain and the peak magnitude and its corresponding frequency in the sliding window are determined. A set of 44 features is extracted from each window and then tested and trained to classify terrain topography using five different machine learning methods: Artificial Neural Network, Decision Tree, k-Nearest Neighbor, Naive-Bayes, and Support Vector Machine. The 44-feature set is optimized using a wrapper selection algorithm for the Decision Tree and k-Nearest Neighbor algorithms. The results show that by utilizing sensor data from an IMU in combination with machine learning methods a terrain topography classification algorithm can accurately predict various terrains over which the user traverses.