Novel Deep convolutional network for elderly activity recognition
Abstract
HAR is an essential factor in elderly care to continuously monitor their daily movements and aid in the case of risks and emergencies. Conventional HAR algorithms tend to miss the complex spatiotemporal relations that describe human motion hence their incapability in the practical world. To overcome this issue, we suggest a new deep convolutional neural network tailored to elderly activity recognition problems. The suggested architecture combines CNN with sophisticated temporal modelling and attention-like techniques to be able to both extract and prioritize important motion features and ignore irrelevant background information. Basing on a benchmark dataset, the model proves to be more accurate in the classification of complex activities with a significant gain in terms of accuracy, precision, recall, and F1-score relative to conventional methods. The obtained results suggest the viability of the suggested framework in coping with a wide range of and noisy settings, which makes the framework a potentially useful solution to the problem of intelligent healthcare monitoring systems. This work has made HAR robust and reliable, therefore contributing to the development of person-tailored, adaptive, and proactive elderly care technologies. However, the time and positional features of the motions are too dissimilar to be captured using a single framework. We thus propose a ConvLSTM3D-HOOF model by effectively combining the strengths of traditional convolution and LSTM spatial-temporal networks. The model performs at the state of the art on the MSR Daily Activity3D dataset (91% accuracy, 91.25% precision, 91.25% recall, and 90.5% F1 score), supporting its performance as an elderly activity detector. We solicit the assistance of a ConvLSTM3D architecture with a new strategy that is Transformer-inspired attention and Adaptive Relevance Attention Module (ARAM) to effectively capture patio-temporal features in a noisy situation. On top of the MSR Daily Activity3D dataset, the model produced a maximum of 94% F1-score and 93% accuracy, which proves its effectiveness in such projects as elderly activity recognition and assisted living.
Keywords
Human Activity Recognition (HAR), Elderly Monitoring, ConvLSTM3D, Histogram of Oriented Optical Flow (HOOF), Adaptive Relevance Attention Module (ARAM), Spatiotemporal Feature Extraction, Deep Learning, Healthcare Support Systems, Fall Detection, Assisted Living Environments.