Page 142 - Kaleidoscope Academic Conference Proceedings 2024
P. 142
2024 ITU Kaleidoscope Academic Conference
passed to the proposed module, the Person localization is Table 1 – Speech classification
achieved with a confidence score of 0.75 and further
Emotion Precision Recall F1-Score
classified as fall detected.
Angry 0.98 0.96 0.97
Disgust 0.96 0.97 0.97
Fear 0.96 0.97 0.96
Happy 0.96 0.95 0.96
Neutral 0.97 0.98 0.97
Sad 0.96 0.97 0.97
Surprise 0.97 0.98 0.98
Figure 3 Realtime testing input image
Figure 5 – Confusion matrix of voice model
Figure 4 - Dataset sample images
Validation accuracy has emerged as a pivotal metric in
The video module combines state-of-the-art techniques in
computer vision and deep learning to monitor the activity assessing the voice model's proficiency. This metric
measured the proportion of correctly classified instances
and posture of elderly individuals in real time. By integrating
MobileNet-based posture detection with attention and within the validation dataset, serving as a reliable indicator
of the overall robustness and reliability of the voice model.
hourglass models, the system ensures accurate and efficient
identification of anomalies, such as falls. This proactive With an impressive validation accuracy of 96.34%, the
proposed system's voice model demonstrated a high level of
approach to elderly care enhances the safety and well-being
of individuals living independently, providing peace of mind accuracy and proficiency in discerning and classifying
different speech patterns and emotions. The relative plot
to both users and their caregivers.
between epochs and training and validation accuracies is
shown in Figure 6.
5. RESULTS
Precision, a fundamental metric in classification tasks,
5.1 Performance of voice-based model
provided valuable insights into the accuracy of the voice
model's predictions. Both weighted average precision and
In the evaluation of the proposed system's voice model
macro average precision were evaluated to assess the model's
performance, several key metrics were meticulously performance across all classes in a multi-class classification
analyzed to ascertain its efficacy in real-world applications.
scenario specific to voice analysis. The weighted average
Among these metrics, the validation loss assumed precision, which accounted for the number of samples in
significance as it quantified the error between the predicted
each class, offered a comprehensive assessment of the voice
labels and the actual labels for the validation dataset, offering
model's overall precision. Similarly, the macro average
insights into the model's generalization capabilities and
precision provided an unbiased measure of the model's
predictive accuracy specific to voice analysis. The system performance across all classes, regardless of class
performance measure on speech classification is listed in
imbalances. With both metrics yielding a consistent
Table-1 and corresponding confusion matrix is depicted in precision score of 96%, the proposed system's voice model
Figure 5.
demonstrated a commendable level of accuracy and
reliability across diverse speech patterns and emotions.
– 98 –