Algorithm 1. Deep learning-based complex activity identification training procedure.

1: Input: Training acceleration sample $X_n$  2: Output: Set of activity detail performances  3: Sensor Data preparation  4: Obtain the acceleration sensor data from smartphone  5: Segment the sensor data using sliding window  6: Compute the magnitude using Equation (16)  7: Compute the pitch-roll values using Equations (17) and (18)  8: Network Parameter Settings  9: Set the number of hidden layers and neurons 10: Max epoch values 11: Sparsity regularization values 12: Train the stacked autoencoder using greedy-wise layer approach 13: Compute the cost function of the autoencoder algorithm at each layer using Equations (3)–(5) 14: Set the sparsity regularization values using Equations (6)–(9) 15: Obtain the network output 16: Stack the pre-trained network with their parameter values 17: Train the Softmax classifier to estimate their parameters 18: Minimize the cost function 19: Fine-tune the stacked autoencoder network weights using gradient descent 20: Obtain the activity details