Analyzing Parameter Patterns in YOLOv5-based Elderly Person Detection Across Variations of Data

This study investigates the impact of data variations on parameter patterns within a YOLOv5-based elderly person detection model. We explore how changes in camera settings and environmental factors influence the model's parameters. Our research questions focus on how these variations affect parameter patterns, model robustness, and generalizability. We aim to identify the parameters and layers of the model most susceptible to variations and develop strategies to improve the model's performance across various datasets. The experiment involves collecting data from eight elderly participants in real-life elder care facility settings. During data acquisition, only depth images are recorded using stereo-depth cameras to protect privacy. After that, we train individual YOLOv5 models for each dataset through hyperparameter tuning and transfer learning. Optimal hyperparameters and the sensitive convolutional layer of each model are then compared. Class Activation Maps (CAMs) are utilized to visualize the network's focus, followed by analysis of weight distributions and correlation to identify parameter patterns. The findings will provide valuable insights for improving elderly person detection models for smart healthcare and their robustness to real-world variations.