Improving Sensor-based Activity Recognition Using Motion Capture as Additional Information

We propose a new method for human activity recognition using a single accelerometer sensor and additional sensors for training. The performance of inertial sensors for complex activities drops considerably compared with simple activities due to inter-class similarities. In such cases deploying more sensors may improve the performance. But such strategy is often not feasible in reality due to costs or privacy concerns among others. In this context, we propose a new method to use additional sensors only in training phase. We introduce the idea of mapping the test data to a codebook created from the additional sensor information. Using the Berkeley MHAD dataset our preliminary results show this worked positively; improving in 10.0% both the average F1-score and the average accuracy. Notably, the improvement for the stand, sit and sit to stand activities was higher, typical activities for which the inertial sensor is less informative when using the wrist-worn accelerometer.