Establishing Pose Based Features Using Histograms for the Detection of Abnormal Infant Movements

Area of Science:

• Biomedical Engineering
• Computer Vision
• Developmental Pediatrics

Background:

• Early diagnosis of cerebral palsy is crucial for intervention.
• General Movements Assessment (GMA) is a promising tool, but objective analysis is needed.
• Automated analysis of infant movement from video can support clinical assessments.

Purpose of the Study:

• To develop and evaluate new pose-based features for infant body movement classification.
• To compare the performance of these features against expert-based GMA.
• To assess the potential for automated early detection of abnormal movements.

Main Methods:

• Extracted 2D skeletal joint locations from RGB images using established methods.
• Developed two novel features: Histograms of Joint Orientation 2D (HOJO2D) and Histograms of Joint Displacement 2D (HOJD2D).
• Trained and tested classifiers (kNN, LDA, Ensemble) using these features on the MINI-RGBD dataset.

Main Results:

• The proposed HOJO2D and HOJD2D features effectively represented infant pose and movement.
• The Ensemble classifier achieved high accuracy (91.67%) in distinguishing normal from abnormal movements.
• Results show promise for objective, automated analysis complementing GMA.

Conclusions:

• Novel pose-based features demonstrate significant potential for automated infant movement analysis.
• This approach can aid in the early and objective diagnosis of conditions like cerebral palsy.
• Further research can refine these features for broader clinical application.