This paper investigates optimal model selectionfor posture recognition. Accuracy and computational timeare related to the trained model in a supervised classification.An optimal model selection is important for a reliableactivity monitoring system. Conventional guidance onmodel training uses large instances of randomly selecteddata in order to characterize the classes. A new approach tothe training of a multiclass support vector machine (SVM)model suited to limited training sets such as used in posturerecognition is provided. This approach picks a smalltraining set from misclassified data to improve an initialmodel in an iterative and incremental fashion. In addition,a two step grid-search algorithm is used for the parameterssetting. The best parameters were chosen according to thetesting accuracy rather than conventional validating accuracy.This new approach for model selection was evaluatedagainst conventional approaches in an activity classificationstudy. Nine everyday postures were classified from abelt-worn smart phone’s accelerometer data. The classificationderived from the small training set and the conventionalrandomly selected training set differed in twoaspects: classification performance to new data (85.1%Pick-out small training set vs. 70.3% conventional largetraining set) and computational efficiency (improved 28%).
|Journal||International Journal of Machine Learning and Cybernetics|
|Publication status||Published (in print/issue) - Mar 2011|
- Optimal model
- Posture recognition
- Multi-class SVM