Training Human Activity Recognition for Labels with Inaccurate Time StampsReviewed
Takamichi Toda, Sozo Inoue, Shota Tanaka, Naonori Ueda,
Ubicomp Workshop for Human Activity Sensing Corpus and its Application (HASCA)
For activity recognition for mobile sensors, we generally use supervised learning when performing activity recognition using mobile sensor devices such as Smartphones. In these methods, training activity levels are required associated with the training sensor data. However, there is risk that the timestamps of the labels are not accurate because this association is done manually with the audio and video that has been acquired along with the sensor information. In this paper, we propose a method of activity recognition that can recognize correct actions even if there are inaccurate time stamps. In this method, we add labels that shift the original training data labels to several possible timestamps. We also implement multi-label machine learning. In addition, we propose a method for repeated training based on the Expectation-Maximization(EM) algorithm. To evaluate this method, we conducted an experiment to recognize three types of behavior using a Naive Bayes classifier. We slid the action labels of the sequence data and examined whether the recognition rate was improved by our proposed method. The results show that the proposed method can perform activity recognition with high accuracy, even if the action labels times are shifted. Using our method can rise the recognition accuracy 48% at the most.