Analisys of the ComplexUpper-LimbMovements dataset with application of some classification algorithms.
The Complex Upper-Limb Movements database contains hand trajectory data collected from ten subjects as they performed various upper-limb motor tasks. The csv data files contain four columns. The data in the first column is the time axis in seconds. The data samples in the other three columns are the x-, y-, and z-coordinates of the reflective marker utilized in each experiment to represent the hand trajectory of movement.
Actions: ['BostonCA', 'BostonCU', 'HarvardCA', 'HarvardCU', 'Can', 'Circle', 'Ellipse', 'Flower', 'Spiral', 'SuperMegaCloud', 'Triangle', 'Planned', 'Unplanned', 'Randomness']
Some actions visualized:
The dataset is divided in windows of lenght TIME_STEPS and with a step of lenght STEP. There are two path followed:
- These windows are feed to some Deep neural networks and
- A feature extraction algorithm is run to extract from the windows some relevant information: [mean, median, std, minv, maxv, percentile25, p50, p75, sum, energy, skewness, kurt]
Different training processes were employed depending on the preprocessing path:
- Different Deep neural network were tested: Fully Connected Network, simple LSTM, bidirectional LSTM, Conv LSTM, ...
- Different non-Deep classifier were tested: Random Forest, SVM, LDA, QDA, ...
The best results were obtained with TIME_STEPS=64, STEP=48.
- For the DNN: the best performances in accuracy were reached with the model Conv LSTM with an accuracy on the test set of: 0.793 (epoch=128. More than 1h to train);
- For the Non-DNN: the best performances in accuracy were reached with the Random Forest Classifier with an accuracy on the test set of: 0.790 (using n_estimators=256, max_depth=None. Some minutes to train the classifier);
Reference: link