Short Segment Random Forest with Post Processing Using Label Constraint for SHL Recognition Challenge
Hitoshi Matsuyama, Kenta Urano, Kei Hiroi, Katsuhiko Kaji, Nobuo Kawaguchi · 2018
The bases of the approaches of UCLab(submission 1) towards SHL recognition challenge are using Random Forest and letting it select important features. Using accelerometer, gyroscope, magnetometer, gravity and pressure sensor as input data, features such as mean, variance, max, difference of max and min, and main frequency are calculated. We find that activities of Still, Train, and Subway are highly similar and hard to distinguish. To achieve robust recognition, we make predictions for every segment of 3 seconds and produce final prediction based on these predictions. Moreover, to deal with the case that one line contains two or more activities, we use a rule-based post processing to predict these activity labels. As a result, using the lines of last 20% in training dataset as validation set, predictions for 3-second segments have around 0.879 of F1-score and predictions for lines have around 0.942.