AI-Powered Vision Sensing System for Real-Time Inspection of Particles in Intravenous Bags Through Multiobject Detection-Tracking Configurations

This research introduces a novel vision-based sensing system for automated inspection of particles in intravenous bags, addressing pharmaceutical quality control challenges. We present a camera-based sensing platform with an Arducam 64 MP autofocus camera (OV64A40 sensor) on a custom electromechanical system, enhanced with a dual-stream transformer architecture that extends detection through continuous spatial-temporal tracking. The vision system shows an enhanced accuracy of 94.3% for detection of microscopic contaminants (0.1-5 mm), with achieved 100% F1-score for critical 0.1-1 mm particles. Transformer-based tracking achieves zero ID switches versus 4-7 in conventional DeepSORT and ByteTrack methods. The optimized architecture processes at 3.53-19.85 FPS, demonstrating feasibility for resource-constrained deployment. This advancement can significantly reduce human error, offer manufacturers flexible high-precision or high-speed configurations, and enhance production efficiency in pharmaceutical quality control.