| dc.description.abstract |
Timely and accurate human fall detection is essential for enhancing the safety
of older adults and individuals who require assistance. In video-based human fall
detection, variations in movement speed and motion patterns across different activities,
as well as changes in body posture and movement throughout an action, can be made
accurate activity recognition challenging. In addition, differences in camera placement,
recording distance, and viewing angles can lead to variations in the appearance and
posture of a person across video frames. Therefore, reliable fall detection requires the
effective analysis of both spatial features related to human appearance and temporal
patterns that represent movement over time. To address these requirements, this
research proposes a deep learning-based video classification model that combines a
Convolutional Neural Network (CNN), a Temporal Convolutional Network (TCN), and
an Attention Mechanism.
The proposed model extracts spatial features related to human posture and
visual appearance from RGB video frames using CNN. The resulting feature sequences
are then processed by TCN to learn movement patterns and temporal dependencies. In
addition, the attention mechanism helps the model emphasize informative features and
temporal segments relevant to fall classification. By integrating spatial and temporal
representations, the model aims to distinguish fall events from post-fall recovery and
other non-fall activities.
The Multiple Cameras Fall Dataset (MCFD) is used for model training and
evaluation. The performance of the proposed CNN-TCN-Attention model was
compared with that of the CNN-LSTM model for fall detection. The result showed that
the CNN-TCN-Attention model achieved a test accuracy of 84.88%, outperforming the
CNN-LSTM model, which obtained 82.56% under the same experimental settings. |
en_US |