• Reduced visibility from cement dust, direct sunlight, and deep shadows • More false detections and misses from unstable exposure
Approach
• Restored clarity with CLAHE and dehaze preprocessing • Applied adaptive correction with dust-sensor thresholds • Improved robustness with hard-negative mining and lighting-based auto-augmentation
Impact
• Mixed-zone detection +36p (48% → 84%) • 0 near-miss incidents over nearly 2 months at Sampyo Cement Plant • 60% fewer false detections in dust and glare zones
High-Angle CCTV & Partial Occlusions
Challenge
• Difficulty distinguishing falls from kneeling or bending in high-angle views • Delayed recognition caused by partial body visibility or equipment occlusion
Approach
• Extracted joint data using multi-frame Skeleton Pose estimation • Trained action recognition models specifically on dynamic fall patterns • Augmented synthetic fall data (varied angles and backgrounds) via simulators
Impact
• Fall detection accuracy +30p (60% → 90%) • 70% fewer false alarms for normal working postures • 95% success rate for instant alerts within 3 seconds of actual falls
Composite PPE Recognition
Challenge
• More false detections from railings, wiring, and other background elements
• Misses caused by small, overlapping PPE items
Approach
• Retrained on 180K labeled site data • Applied Mosaic/CutMix for small-object enhancement • Fine-tuned YOLOv8 with Focal Loss to reduce misses
Impact
• PPE mAP +34p (54% → 88%) • 91% accuracy for no-PPE alerts • 80% fewer false positives and misses per hour (10 → 2)
Night Flare
Light Bleed
Challenge
• Overexposure and flare from direct lighting • More missed detections from low nighttime contrast
Approach
• Reduced flare with IR coaxial lighting and a 15° camera adjustment • Applied HDR, tone mapping, and nighttime flare fine-tuning
• Reduced visibility from cement dust, direct sunlight, and deep shadows • More false detections and misses from unstable exposure
Approach
• Restored clarity with CLAHE and dehaze preprocessing • Applied adaptive correction with dust-sensor thresholds • Improved robustness with hard-negative mining and lighting-based auto-augmentation
Impact
• Mixed-zone detection +36p (48% → 84%) • 0 near-miss incidents over nearly 2 months at Sampyo Cement Plant • 60% fewer false detections in dust and glare zones
High-Angle CCTV & Partial Occlusions
Challenge
• Difficulty distinguishing falls from kneeling or bending in high-angle views • Delayed recognition caused by partial body visibility or equipment occlusion
Approach
• Extracted joint data using multi-frame Skeleton Pose estimation • Trained action recognition models specifically on dynamic fall patterns • Augmented synthetic fall data (varied angles and backgrounds) via simulators
Impact
• Fall detection accuracy +30p (60% → 90%) • 70% fewer false alarms for normal working postures • 95% success rate for instant alerts within 3 seconds of actual falls
