Deep Custom Transfer Learning Models for Recognizing Human Activities via Video Surveillance

Abstract

The use of video surveillance for human activity recognition (HAR) in inpatient rehabilitation, activity recognition, or mobile health monitoring has grown in popularity recently. Before using it on new users, a HAR classifier is often trained offline with known users. If the activity patterns of new users differ from those in the training data, the accuracy of this method for them can be subpar. Because of the high cost of computing and the lengthy training period for new users, it is impractical to start from scratch when building mobile applications. The 2DCNNLSTM, Transfer 2DCNNLSTM, LRCN, or Transfer LRCN were proposed in this paper as deep learning and transfer learning models for recognizing human activities via video surveillance. The Transfer LRCN scored 100 for Training Accuracy and 69.39 for Validation Accuracy, respectively. The lowest Validation Loss of 0.16 and the Lowest Training Loss of 0.001 was obtained by Transfer LRCN, respectively. The 2DCNNLSTM has a 98.34 lowest training accuracy and a 47.62 lowest validation accuracy.