A Transferred Daily Activity Recognition Method Based on Sensor Sequences

The feature-based transfer learning method has become popular for transferred daily activity recognition in heterogeneous smart home environment since the feature-based transfer learning can reduce the difference between the source smart home environment and the target smart home environment. However, feature-based transfer learning has a poor recognition effect on daily activity recognition, for example, when several daily activities with similar features are transferred by feature similarity, the recognition results may be wrong. To improve recognition performance between similar daily activities, this paper presents a novel daily activity recognition method based on sensor sequence similarity and a global similarity calculation method based on random sampling. Firstly, the sensor similarity between source domain and target domain will be calculated to generate the sensor mapping matrix. Secondly, the features are extracted from all the samples of daily activities to calculate the similarities. Thirdly, the sensor similarity matrix is employed to calculate the sensor sequence similarity between the samples. Finally, the source domain samples with high similarity are chosen as similar samples to calculate global similarities with a random sampling method. The best matching source sample will be selected and the label is sent to the target sample. To evaluate our method, we use four public datasets to prove the validity of our method. The result of the experiment shows that our method is better than other commonly used methods.