Hybrid CNN-SVM Approach for Time-Series Activity Recognition Using Wearable Sensors

Wearable electronics generate copious amounts of time-series data that are highly valuable in understanding user activity and physiological state. However, the use of this data for accurate activity classification poses a challenge thanks to issues of dimensionality, noise, and temporal correlation. This work proposes a combined pipeline for the FVC2006 databases using CNN for feature extraction and SVM for classification to overcome these challenges. Using the UCI Human Activity Recognition dataset of accelerometer and gyroscope data with six groups of activities we tested the proposed methodology. The CNN-SVM pipeline achieved a higher accuracy of 92.5 % in comparison to CNN (88.4 %) individually and SVM (79.2%). The classifications within all the activity classes were well-done with statistically similar scores of precision at 91.8%, recall at 92.0%, and F1-score at 91.9%. Misclassifications for comparable activities included “Upstairs” versus “Downstairs,” a misclassification rate of 4.5 percent, and “Standing” versus “Lying,” a misclassification rate of 3.8 percent. The resultant feature extraction was shown with the help of t-SNE map where inter class distance was found to be 88.5% and intra class distance was only 12.7%. From the earlier training and validation curves we see that both these curves gave us good indication of the model converging and validation accuracy of 92.3%. In fact, this paper presents the CNN-SVM as a highly efficient approach to dealing with certain related intricacies involved in time-series data for activity recognition in healthcare and fitness monitoring. Future prospects will focus on online implementation and learning technologies for different sorts of data.