Enhancing behavior classification of children in dynamic interaction scenes through improved DCNN model.

Area of Science:

• Child Development
• Computer Vision
• Environmental Science

Background:

• Societal development increases focus on children's environmental quality.
• Individualized environments are crucial for children's growth due to developmental differences.
• Dynamic interaction scenarios necessitate adaptive environmental solutions.

Purpose of the Study:

• To propose an environment creation method for children's behavior classification.
• To enhance the quality of children's growth environments through tailored settings.
• To bridge the semantic gap between environmental and behavioral features.

Main Methods:

• Utilizing an encoder-decoder architecture to classify children's behavior from video data.
• Employing a Deep Convolutional Neural Network (DCNN) backbone for feature extraction (shallow and high-level features).
• Leveraging the DenseNet model to minimize the semantic gap and maximize feature similarity between behavior and environment.

Main Results:

• Accurate classification of children's behavior with an F-score exceeding 70%.
• Successful feature fusion using DenseNet's dense blocks for multi-modal similarity calculation.
• Demonstrated capability to identify suitable environments based on behavior characteristics.

Conclusions:

• The proposed method accurately classifies children's behavior, providing a foundation for environment creation.
• This approach ensures appropriate environmental conditions, supporting healthy child development.
• The model offers a scientific basis for guaranteeing optimal growth environments for children.