An Improved Empirical Hyper-Parameter Tuned Supervised Model for Human Activity Recognition based on Motion Flow and Deep Learning

摘要

Traditional pattern recognition methods rely on manual feature-extraction, which may result in the poor generalization of the model. With the increase in the popularity and success of deep learning methods, it is widely adopted in Human Activity Recognition (HAR). The ability of HAR can be extended to automated surveillance systems. In this paper, a deep learning and motion flow based Incept_LSTM is proposed. The proposed method extends the capability of pre-trained Inception-v3 and Long Short-Term Memory (LSTM). The hybridization of these models sustains a spatio-temporal convergence which is validated by the results so obtained. The proposed model is trained and validated on UCF-Crime dataset. The obtained results are then compared with the work done in the literature on the UCF-Crime dataset, KTH, and UCF-Crime2Local. It has achieved an accuracy of 98.2% and 94.57% on training and validation, respectively. Testing the effectiveness of RMSProp optimizer (as opposed to Adam) with 1e-6 learning rate has given best fit with 0.2 training and 0.38 validation loss. The model takes the advantage of motion flow computed using Lucas-Kanade Method. Motion flow is the important paradigm for considering video data. The proposed method outperforms the state-of-the-art methods in terms of accuracy, number of parameters and processing time. Also, various hyper-parameter settings are performed for the best training results.