Early autism detection is hindered by subjective assessments, delays, and limited resources, highlighting the need for automated tools. This study introduces a multi-modal video analysis system combining skeletal stickman, YOLOv8 pose estimation, motion tracking, and facial analysis to build behavioural profiles. Using Random Forest on movement features (stereotypies, asymmetry, temporal smoothness) with spectral analysis (0.3–7 Hz), the system exceeds 90% accuracy on the ComplexVideos dataset (100 subjects) with real-time performance (30 FPS). Feature importance shows movement variability and temporal dynamics as key discriminators. Future directions include clinical validation, biomarker integration, and age-specific models to improve diagnostic precision.