Table 4.

Evaluation of representative inertial sensor-based gait recognition approaches (part 2/2).

Approach	Experiment Description					Length of Shortest Gait Epoch Used for Recognition	Validation		Performance		Special Remarks
	Dataset Reference	Type of Validation	No. Subjects (M + F)	Protocol Description	Measurement Length		Gallery Data	Probe Data	Measure	Value
Ngo et al., 2014 [28]	[28]	Experimental	47 (32 + 15)	16 trials per subject: two days, 2 weights, 4 sensors	Each trial 2 min, about 64 gait periods, 90 m long walking path		Data acquired on first day (by 3DM-GX3-25 sensor)	Data acquired on second day (from all sensors)	EER	10%
B. Sun et al., 2014 [53]	[53]	Partly realistic	10	Straight walk on two surfaces: pavement and grass, 40 sets of data for each subject	Each trial 10 s (9–10 gait cycles)	10 s	One set of data for one subject	Remaining 3 sets of data	Accuracy	All correct
Ren et al., 2014 [52]	[52]	Realistic	26	Casual walking of users, 3048 trials in half year, 2 types of trials: short and long; experiment included gait speed variations as well as spoofing scenario (8 adversary and 10 spoofing users)	Long trials: about 10 min; short trials: 10, 20 and 40 s (detection latency, walking speed and placement studies)	20 s for stable accuracy	Several gallery and probe pools for different evaluation phases		Accuracy, FRR	Accuracy over 80% (user-side), over 90% (server side), FP rate under 10%	Includes important studies: step cycle identification, detection latency, walking speed, placement and possibility of spoofing
Zhang et al., 2014 [29]	[29]	Experimental	175 (153 in seasons S1 and S2, 22 in one season S0)	2 recording seasons on level walk, 6 trials per subject in one season, 1 week–0.5 year time interval between two seasons	20 m straight level walk, 7–15 s for single trial (7-14 gait cycles)	7–15 s	Identification: S1 or S2 for enrolment (as well as S0), remaining for identification; authentcation: S1 and S2 into threefolds, multiple targets per fold and probes per target (exhaustive protocol		EER (authentication), accuracy (identification)	95.8% accuracy for identification, 2.2% EER for authentication	Exhaustive evaluation, data acquired from multiple sensors simultaneously
Zhong et al., 2014 [45]	[27,44]	Experimental ([27]), realistic ([44])	*	*	*	Entire signals	*	*	EER (experimental), accuracy (realistic)	Experimental: 6.8% EER (accelerometer), 10.9% EER (gyrometer), 5.6% EER (fused); realistic: 66.3% accuracy
Hoang et al., 2015 [54]	[78]	Partly realistic	38 (28 + 10)	Acquisition of 16 gait templates, each gait template consists of 4 consecutive gait cycles	At least 64 steps to generate 16 gait templates	8 random gait templates	Half-half random selection of gait templates		EER, FAR, FRR	0%, 16.2%, 3.5%
Sprager et al., 2015 [46]	[27,44]	Experimental ([27]), realistic ([44])	*	*	*	1.4 s (both experimental cases), 12 s (realistic)	*	*	EER (experimental), accuracy (realistic)	Experimental, single sensor: 10.1% EER, sensor fusion: 5.5% EER; realistic: 69.4% accuracy	Experiment on very short gait epochs, variable epoch length