**5. Evaluation Results**

We first give a figurative example to present the WD. We apply THR and SVM walk detection algorithms to the continuous activities in a real indoor scenario. The activities consist of a walk, using a phone, going up stairs, going down stairs and some temporal irregular activities such as pushing the door or handshake.

The results are shown in Figure 4. The windows surrounded by red and green rectangles indicate the walking state recognized from SVM and THR, while the low values out of the rectangles are non-walk activities. We saw that SVM provides better classification accuracy than the threshold method in this example. More details will be shown in the following experiments. We could find there are some jitters along the timeline in both algorithms; in addition, the estimated start and end times of walking may deviate from the ground truth.

**Figure 4.** Example of walk detection in a real timeline.

We have defined four groups: walk, walk-like activity, walk-related activity and walk-unrelated activity. We first address a coarse-grained WD problem that distinguishes walk and walk-like activity from walk-related activity and walk-unrelated activities.
