Human Activity Recognition Using Tag-Based Radio Frequency Localization

Figures & data

Figure 1. (a) Xsens MTx (Xsens 2016) and (b) 3DM-GX2 (MicroStrain 2016) sensor units.

Figure 2. (a) Motion sensor units worn on the body (Xsens 2016) © Xsens. Reproduced by permission of Xsens. Permission to reuse must be obtained from the rightsholder. (b) An active RFID tag (SYRIS SYSTAG245-TM-B) worn as a bracelet (SYRIS 2016) © SYRIS. Reproduced by permission of SYRIS. Permission to reuse must be obtained from the rightsholder. (c) An RFID tag inserted under the skin © GeekWire.com. Reproduced by permission of GeekWire.com. Permission to reuse must be obtained from the rightsholder. (d) Tiny RFID tags of size 2mm x 2mm © Tagent Corp. Reproduced by permission of Tagent Corp. Permission to reuse must be obtained from the rightsholder.

Figure 3. Ubisense hardware components (Steggles and Gschwind 2005). © Ubisense. Reproduced by permission of Ubisense. Permission to reuse must be obtained from the rightsholder.

Figure 4. Sample rows of the original dataset.

Figure 5. The positions of (a) tag 1 and (b) tag 3 in the first experiment of the first subject as 3D curves whose gray level changes from light gray to black over time.

Figure 6. The three curve-fitting methods applied to synthetic position data.

Figure 7. The x position of tag 4 in the fifth experiment of the fifth subject. (a) The whole curve and (b) the zoomed-in version.

Table 1. The average probability of error (weighted by prior probabilities) of the 10 classifiers, for the 11- and 5-class problems and the two cross-validation techniques. In each case, the combination of parameters leading to the most accurate classifier is given in parentheses.

Classifier		Error Percentage ± Standard Deviation (with the Corresponding Parameter Values fs, frm_dur, curve_fit_type, reduc)*
		11 Classes		5 Classes
		5-Fold	Subject-Based L1O	5-Fold	Subject-Based L1O
1	Gaussian classifier with the same arbitrary covariance matrix for each class	20.89 ±  0.02 (10, 0.5, 1, −8)	23.52 (10, 0.5, 3, −8)	6.06 ±  0.01 (10, 0.5, 1, −4)	7.63 (10, 0.5, 1, −4)
2	Gaussian classifier with different arbitrary covariance matrices for each class	20.76 ±  0.04 (10, 0.5, 1, −6)	23.69 (10, 0.5, 3, −8)	5.48 ±  0.02 (10, 0.5, 1, −4)	7.73 (10, 0.5, 1, −4)
3	Gaussian classifier with different diagonal covariance matrices for each class	21.65 ±  0.05 (10, 0.5, 1, −6)	24.11 (10, 0.5, 3, −3)	6.02 ±  0.02 (10, 0.5, 1, −4)	7.92 (10, 0.5, 1, −4)
4	mixture of Gaussians classifier (with two mixtures)	20.75 ±  0.07 (10, 0.5, 1, −6)	23.70 (10, 0.5, 3, −8)	5.19 ±  0.03 (10, 0.5, 1, −4)	7.63 (10, 0.5, 1, −4)
5	naïve Bayes classifier	22.84 ±  0.16 (10, 0.5, 1, −8)	24.46 (10, 0.5, 3, −6)	7.24 ±  0.16 (10, 0.5, 1, −4)	9.49 (10, 0.5, 1, −4)
6	k-NN classifier (k = 5)	8.67 ±  0.10 (10, 0.5, 3, 0)	21.30 (10, 0.5, 1, −10)	1.12 ±  0.04 (10, 0.2, 1, 0)	6.52 (10, 0.2, 1, 0)
7	dissimilarity-based classifier	19.41 ±  0.16 (10, 0.5, 1, −8)	22.10 ±  0.04 (10, 0.5, 3, −10)	4.66 ±  0.16 (10, 0.5, 1, −4)	7.33 ±  0.03 (10, 0.5, 1, −4)
8	minimum least-squares linear classifier	25.26 ±  0.52 (10, 0.5, 3, −10)	26.83 (10, 1, 3, 0)	7.37 ±  0.17 (10, 0.5, 1, −4)	9.30 (10, 1, 1, 0)
9	nearest mean classifier	23.60 ±  0.06 (10, 0.5, 3, −10)	27.02 (10, 0.5, 3, −10)	6.07 ±  0.02 (10, 0.5, 1, −4)	8.21 (10, 0.5, 1, −4)
10	scaled nearest mean classifier	20.94 ±  0.04 (10, 0.5, 3, −8)	23.59 (10, 0.5, 3, −8)	6.04 ±  0.03 (10, 0.5, 1, −4)	7.58 (10, 0.5, 1, −4)

^*pri = 1 throughout the table; that is, prior probabilities are calculated from the class frequencies.

Table 2. Cumulative confusion matrices for classifier 6 (k-NN) for the 11-class problem. The confusion matrices are summed up for the five executions of the 5-fold (top) and subject-based L1O (bottom) cross validation.

Cumulative Confusion Matrix of Classifier 6 (k-NN) for 5-fold (Average Classification Error: 8.67%) (fs = 10, frm_dur = 0.5, curve_fit_type = 3, pri = 1, reduc = 0)
True Labels	Estimated Labels											Total
	1	2	3	4	5	6	7	8	9	10	11
1	10,057	177	158	171	113	45	298	45	23	56	52	11,195
2	139	606	18	124	16	94	42	0	62	16	3	1,120
3	139	14	1,752	146	3	6	64	65	30	1	15	2,235
4	126	108	131	17,556	14	7	310	45	6	6	11	18,320
5	124	34	6	30	321	96	22	0	4	13	0	650
6	41	59	6	19	60	8,758	3	0	20	67	7	9,040
7	305	45	54	305	10	13	5,691	5	5	1	11	6,445
8	30	1	52	76	0	0	7	1,612	0	0	2	1,780
9	16	42	26	5	2	9	9	0	3,729	0	52	3,890
10	74	19	0	12	9	54	10	0	0	332	0	510
11	50	2	9	11	0	5	3	0	54	0	846	980
Total	11,101	1,107	2,212	18,455	548	9,087	6,495	1,772	3,933	492	999	56,165
Cumulative Confusion Matrix of Classifier 6 (k-NN) for Subject-Based L1O (Average Classification Error: 21.30%) (fs = 10, frm_dur = 0.5, curve_fit_type = 1, pri = 1, reduc = − 10)
True Labels	Estimated Labels											Total
	1	2	3	4	5	6	7	8	9	10	11
1	10,645	30	65	15	30	195	190	0	5	10	10	11,195
2	305	340	10	135	10	115	170	5	25	0	5	1,120
3	215	20	510	415	15	80	850	85	0	10	35	2,235
4	40	100	220	16,665	0	30	575	350	285	0	55	18,320
5	180	20	25	5	110	190	95	0	0	25	0	650
6	310	30	45	20	85	8,390	135	0	10	0	15	9,040
7	685	110	355	1,385	50	235	3,110	215	75	25	200	6,445
8	5	10	50	625	5	20	370	630	10	0	55	1,780
9	0	50	0	135	5	175	60	0	3,380	0	85	3,890
10	180	5	30	0	40	150	80	0	5	20	0	510
11	110	0	0	25	20	60	200	5	140	0	420	980
Total	12,675	715	1,310	19,425	370	9,640	5,835	1,290	3,935	90	880	56,165

Table 3. Cumulative confusion matrices for classifier 6 (k-NN) for the 5-class problem. The confusion matrices are summed up for the five executions of the 5-fold (left) and subject-based L1O (right) cross validation.

Cumulative Confusion Matrices of Classifier 6 (k-NN) (fs = 10, frm_dur = 0.2, curve_fit_type = 1, pri = 1, reduc = 0)
True Labels	5-Fold (Average Classification Error: 1.12%)						Subject-Based L1O (Average Classification Error: 6.52%)
	Estimated Labels					Total	Estimated Labels					Total
	1	4	6	8	9		1	4	6	8	9
1	27,041	10	226	7	6	27,290	26,850	15	350	50	50	27,290
4	39	45,051	15	206	4	45,315	405	42,905	470	1,040	495	45,315
6	145	3	22,296	10	11	22,465	1,050	25	21,030	285	75	22,465
8	59	361	74	3,885	1	4,380	270	735	795	2,540	40	4,380
9	10	0	29	0	9,591	9,630	90	75	775	0	8,690	9,630
total	27,294	45,425	22,640	4,108	9,613	109,080	28,665	43,755	23,420	3,890	9,350	109,080

Figure 8. Effect of sampling frequency on the average classification error of the k-NN classifier (classifier 6).

Figure 9. Effect of segment duration on the average classification error of the k-NN classifier (classifier 6).

Figure 10. Effect of curve-fitting method on the average classification error of the k-NN classifier (classifier 6).

Figure 11. Effect of prior probabilities on the average classification error of the k-NN classifier (classifier 6).

Figure 12. Effect of feature reduction with (a) PCA and (b) LDA on the average classification error of the k-NN classifier (classifier 6).

Xsens Technologies B. V. 2016. MTi and MTx user manual and technical documentation Enschede, The Netherlands: Xsens http://www.xsens.com. (accessed February 2016).

 Google Scholar

MicroStrain. 2016. MicroStrain inclinometers and orientation sensors. Williston, VT USA: MicroStrain, http://www.microstrain.com/3dm-gx2.aspx ( accessed February 2016).

 Google Scholar

SYRIS Technology Corp. (The RFID Total Solution). 2016. Address: 12F., No.16, Sec. 2, Taiwan Blvd., West Dist., Taichung City 403, Taiwan. http://www.syris.com.

 Google Scholar

Steggles, P., and S. Gschwind. 2005. The Ubisense smart space platform, Technical Report, Ubisense, Cambridge, UK, http://www.ubisense.net/. (accessed February 2016).

 Google Scholar