TabIe III: TA accuracies for all scenarios using Dataset 1.
4.2 Model Performance Based on Data Pre-processing Approach For this dataset we made an initial experiment using the TA approach in which the results were well below those from CASAS experiments. They used a baseline of three different models, Naive Bayes, HMM and CRF, getting accuracies of 80.33 %, 75.55 % and 90.76 % respectively. The accuracy for our SVM combined with the TA approach was of only 51 %, which indicated that the TA preprocessing was not the appropriate choice for this dataset. Following the procedures used by the CASAS researchers, we applied the CDA preprocessing to the dataset, and the accuracy of our SVM model increased to 91.77 %, better than any other result so far.
Table IV: The results using Dataset 2 and TA approach showed a poor performance. However, the same data combined with the CDA approach, showed high performance specially when using discriminative models. The multi-class SVM model achieved the best performance of 91.67 %.
As happened with Dataset 1 results, SVM again outperforms all other models considered for CDA processing approach, that is, both models we proposed( HMM and KNN) and the results obtained by CASAS algorithm which best result was 90.77 % with a CRF approach. Although when using the TA( 60 seconds timeslice) approach HMM performs better( 51.60 % vs. 59.85 %) than SVM, it is clear that, regardless the algorithm used, this TA is not the best approach to process this data, which is clearly meant to be processed using CDA.
5 Conclusion
In this work, we have presented an activity pattern recognition model using a non-linear multi-class SVM approach for detecting daily activities based on two public available datasets. We have compared our methods with other state of the art machine learning approaches using the same datasets. The result demonstrates the proposed method outperforms other methods. We have developed two data preprocessing techniques including TA( time slice) and CDA( chunk data). It is noted that appropriate data preprocessing techniques can significantly improve the accuracy of the model. Future work will be to test the models using a wider variety of datasets, data processing approaches and new mathematical modelling approaches to establish more comprehensive model performance baselines.
252 ZEMCH 2015 | International Conference | Bari- Lecce, Italy