Table 1.

Summary of human activity datasets used to evaluate the proposed deep learning models. Training window length indicates the number of samples in a window that we found to yield the best results for each dataset. Each dataset was divided into 80% for training and 20% for testing.

Dataset	# of Classes	Sensors	# of Subjects	Sampling Rate	Training Window Length	# of Training Examples	# of Testing Examples
UCI-HAD [19]	6	3D Acc., Gyro., and Magn. of a smartphone	30	50 Hz	128	11,988	2997
USC-HAD [20]	12	3D Acc. & Gyro	14 (5 sessions)	100 Hz	128	44,000	11,000
Opportunity [21]	18	7 IMU sensors (3D ACC, Gyro & Mag.) & 12 Acc.	4 (5 sessions)	30 Hz	24	55,576	13,894
Daphnet FOG [22]	2	3 3D Acc.	10	64 Hz	32	57,012	14,253
Skoda [23]	11	3D Acc.	1 (19 sessions)	98 Hz	128	4411	1102