Improving the Accuracy of Travel Mode Detection for Low Data Collection Frequencies

Abstract

Smartphones are a necessary part of our daily lives, and are equipped with a range of sensors. These two characteristics make them a much preferred data collection device, when compared with other wearable sensors and devices. However, due to huge amount of data collected, issues like processing cost and battery consumption are limiting factors. In order to tackle these issues, the current study aims at achieving acceptable detection accuracy while decreasing the data collection frequency, hence reducing processing cost and battery consumption. To begin with, several classification algorithms are compared. Results suggest that boosted decision tree provides the highest accuracy closely followed by random forest; however, random forest is preferred because it requires less processing time. Detection accuracies are calculated at various data collection frequencies, and subsequently improved by addressing the issue of imbalanced data with the introduction of weighted random forest. Further improvement is achieved by applying a two-step post-processing method. Overall accuracy for 0.2 Hz frequency data is improved from 94.98% to 98.78%, whereas for 0.067 Hz frequency, the increase is from 89.16% to 95.40%. Accuracy drop of 3.42% from 0.2 Hz to 0.067 Hz is tolerable because it results in 81.96% decrease in processing time.