Evaluation of supervised classification algorithms for human activity recognition with inertial sensors

The main aim of this work is to compare the performance of different algorithms for human activity recognition by extracting various statistical time domain and frequency domain features from the inertial sensor data. Our results show that Support Vector Machines with quadratic kernel classifier (accuracy: 93.5%) and Ensemble classifier with bagging and boosting (accuracy: 94.6%) outperforms other known activity classification algorithms. A parallel coordinate plot based on visualization of features is used to identify useful features or predictors for separating classes. This enabled exclusion of features that contribute least to classification accuracy in a multi-sensor system (five in our case), made the classifier lightweight in terms of number of useful features, training time and computational load and lends itself to real-time implementation.