Sliding variance and data range for lightweight sports activity recognition with fusion of modalities

This research develops a novel lightweight sport activity detection system using time series data of the accelerometer sensors embedded in two modalities, the smartphone and smartwatch. This research focuses on the lightweight exercises, more specifically, jumping jack, push up, sit up, and squat jump. Such activities are chosen for two reasons; the smartphone is now accessible for many persons, and the lightweight exercises are deemed to be easily completed in daily basis for everyone. Our proposed algorithm includes two novel feature extraction methods, i.e. sliding variance and data range, combined with a digital filter and data clipping methods. The results of feature extraction processes were classified using a combination of k-NN and DTW algorithms. The classification results are subsequently compared with the-state-of-the-art algorithms, i.e. LMNN and Naive Bayes algorithms. The final results imply the merge of k-NN and DTW algorithms (k = 1) with data range method achieves the highest accuracy. The average accuracy for this method is 97.4%, with the processing time of 0.86 s. Thus, the result of counting activity method was acceptable with average values 80% for the whole movement by using two sensor accelerometers.