5.3.2. Baseline Performance

The baseline performance of SC algorithms is obtained under the sampling rate *R* = 200 Hz, orientation *IO* is unknown, personalization *IP* is unknown, and *L* is in Group I or II, where Group I contains Foot, FrontPocket, BackPocket and Hand and Group II contains Hand and UpPocket. The baseline performance of Group I is in Figures 9 and 10; the baseline performance of Group II is in Table 7. For Group I, we could observe one period in one gait cycle of the leg that has a sensor; while for Group II, we could observe two periods.

**Figure 9.** ROC of step counting (Group I).

Figure 9 shows the baseline performance of the algorithms under Group I. PTM, STFT and FSM have higher TPR when FPR is 5%. The overall performance of STFT is best, and it keeps improving along with the increase of FPR. FSM has the highest accuracy, while it is not robust, since it performs poorly when the FPR is low; however, this means that one could obtain better accuracy by fine-tuning the parameters. DWT2 performs poor in this case, because the details are not stable features in the signal of a gait cycle. Although PTM includes a series of elegant signal processing modules, the TPR is good only in a short interval (*f pr* ≈ 4.5%).

**Figure 10.** Error sources distribution (fp ≈ 0.05).

Figure 10 reveals the error proportion of each placement that accounts for the total of 5% false positives (approximately). PTM is the most stable algorithm, and most false positives of STFT and FSM happen at Hand.

Table 7 presents the baseline performance of the algorithms under Group II. We could see that these algorithms outperform Group I remarkably, mainly because the features in each gait cycle are consistent under different contexts such as placements and subjects.


**Table 7.** Performance of step counting (Group II).
