8.2.3. Classification Performance

Figures 10 and 11 summarise the classification performance of the proposed motion mode classifier for detecting each motion mode. As illustrated in Figure 10, the columns of the confusion matrix refer to the ground truth motion modes performed by the participants, and the rows refer to the motion modes predicted by the classifier. The percentage of prediction accuracy, together with their actual number for each motion mode, is presented along the principal diagonal in black colour. The percentage of confused classification for the motion modes are reported along the off-diagonal sections in white colour.

As can be seen from the confusion matrix, the classifier can detect the correct motion mode in more than 95% of cases, irrespective of the type of motion class performed by a participant. The highest accuracy of 99.5% was attained by the classifier for the static motion mode. As reported, 2 and 3 segments out of 928 segments for static type motions were misclassified as striding motion and intermittent motion, respectively. The main reason behind this confusion is because of the simulated door opening activities performed by the participants, which are actually considered as static activities; however, this type of activity sometimes may generate high energy and variance in the signal which can satisfy the decision thresholds, leading to misclassification.

**Figure 10.** Confusion matrix of the proposed motion mode classifier.

**Figure 11.** Classification performance of the proposed motion mode classifier for each motion mode.

The lowest accuracy of 95.4% was achieved for detecting the intermittent motion mode by the classifier. This is because intermittent motions are more likely to be confused as other types of motions. As reported in the confusion matrix, among the total 696 segments for intermittent motions, 17 were detected as striding motion. One possible reason for this confusion may be the consecutive occurrence of a similar type of movement several times (e.g., multiple bending activities), which can produce periodicity and lead to that misclassification. However, during the landmark identification phase, it is very unlikely that such intermittent movement will occur in a pattern that can satisfy the landmark rules. Thus, the defined rules for landmark identification can mitigate the impact of this type of misclassification. Moreover, 17 segments out of 1020 for the striding motion mode were misclassified as intermittent motion, causing the accuracy for that class to be 96.8%.

As presented in Figure 11, the overall sensitivity, specificity, precision, and F-measure of the proposed motion mode classifiers are 97.03%, 99.28%, 95.17%, and 96.06%, respectively, which can eventually produce high accuracy for the landmark identification task.

#### *8.3. Evaluation of User's Body Shadowing Effect Compensation Model*

The proposed BSE compensation model was evaluated firstly using the same experimental data as shown in Figure 2. The polar plot for the raw data is shown in Figure 12, where the distance between the sender and the receiver is 9.9 m. It is noticeable from the figure that the minimum and maximum attenuation of RF signals occur at the angle

positions as discussed in Section 3.2, due to the effect of the user's body. The higher RSSI value is observed at approximately 0◦ with LOS angle positions, and the lowest values of RSSI can be observed at approximately 90◦ and 270◦ angle positions for the raw RSSI data.

**Figure 12.** User's body effect on RSSI (raw RSSI vs. orientation angle).

As the human body shape is uneven, the amount of RSSI depletion, while facing the different body parts, by the signal will be different. The intensification parameter *σ* boosts the RSSI with an amount that can cope with the loss caused by the user's body.Using a single value of *σ* for every orientation angle may lead to an overestimation or an underestimation for some orientations. Thus, three different *σ* values were chosen based on the analysis, as described in Section 3. To evaluate the compensation model performance, firstly the values for the intensification parameters *σ* were empirically investigated, and optimal values were selected, as described in Section 6.2. Figures 13 and 14 present the results after applying the proposed compensation model for two different combinations of values for *σ*. To show the effect of *σ* on the proposed compensation model, first the values of *σ*1, *σ*2, and *σ*3 were chosen as 0.1, 4.0, and 3.0, respectively. Figure 13 shows the polar plot for the compensated RSSI after applying those values to corresponding orientation angles.As can be noticed from the figure, though the model can correct some RSSI values by applying this set of *σ* values that are affected by the user's body shadowing, the corrected values are the underestimation of the 0◦ faced LOS value. Figure 14 presents the results with the values of *σ*1, *σ*2, and *σ*3 as 0.1, 2.0, and 1.5, respectively, which were found as the optimal values for the experimented indoor environments. As can be seen from that figure, the proposed model can considerably correct the attenuated RSSI values while presenting a negligible amount of noise. Most of the corrected RSSI values are almost similar to the 0◦ faced LOS value with a negligible deviation. Thus, the values of *σ*1, *σ*2, and *σ*3 were chosen as 0.1, 2.0, and 1.5 to compensate the user's body affected RSSI values for the proposed indoor localisation system.

**Figure 13.** Results after applying the body shadowing effect compensation model with *σ*1 = 0.1, *σ*2 = 4.0, *σ*3 = 3.0, and *d* = 25.

**Figure 14.** Results after applying the body shadowing effect compensation model with *σ*1 = 0.1, *σ*2 = 2.0, *σ*3 = 1.5, and *d* = 25.

To validate the efficiency of the proposed model, RSSI data were also collected using the same experimental setup, except the wearable device was placed on the back of a user. Figure 15 illustrates the compensation results along with the raw RSSI for the two different sets of *σ* values. As can be seen from the figure, the values of *σ*1, *σ*2, and *σ*3 as 0.1, 2.0, and 1.5 can produce the best estimates for most of the angle positions, compared to the other set.

**Figure 15.** Raw and corrected RSSI for data collected with back-mounted wearable device.

The impacts of the parameter *d*, which represents the distance between the wearable tag and the edge of the far-ended shoulder, were also examined. Figure 16 presents the results of the proposed model for different *d* values, including 41, 25, 50, and 20. Considering the average shoulder width of 41 cm, selecting the value of *d* as 25 produces the best results demonstrated in the figure.

**Figure 16.** Results after applying different values of *d* with *σ*1 = 0.1, *σ*2 = 2.0, *σ*3 = 1.5.

*8.4. Evaluation of the Proposed Localisation System*

To evaluate the performance of the proposed localisation system, experiments were performed in an office building. The experiment building was Building 72 at the Clayton campus of Monash University, which comprises three floors. The experiment was carried out on the second floor, with a total area of about 1475 m2, and the length of the test path was about 150 m, represented by the green line in Figure 17. The path starts from the red circle, then follows the path, and ends at the red square.

**Figure 17.** Experimental area with the markings of experimented path and distribution of reference nodes.
