Sistem Pencatatan Barang Masuk Gudang Berbasis Identifikasi Frekuensi Radio dan Pengenalan Visual Kondisional

Abstract

The process of recording incoming goods at the Multimart Tomohon warehouse is currently entirely manual using paper and spreadsheets, resulting in four systemic problems: data entry errors due to human fatigue, delays in stock updates that hinder real-time decision-making, the inability to reliably track item history, and the absence of a cross-verification mechanism between supplier-claimed quantities and physical quantities received. This research aims to design and test a prototype of an Internet of Things-based incoming goods recording system that integrates radio frequency identification (RFID), infrared sensors, Optical Character Recognition (OCR), and a Convolutional Neural Network (CNN) in a two-box architecture connected by a belt conveyor. The development methodology follows the SDLC Prototyping model consisting of five iterative stages: requirements gathering, rapid design, prototype construction, evaluation and testing, and refinement. The MobileNetV2 architecture was chosen as the CNN backbone because it uses depthwise separable convolution that produces a lightweight model with competitive accuracy, suitable for servers that also process RFID data simultaneously. RFID testing demonstrated accurate tag reading at distances of 1-4 cm with slow to moderate conveyor speeds (∼3-8 cm/s), while at distances of 5-7 cm or high speeds (∼15 cm/s) readings became unreliable. Training a MobileNetV2 CNN model with 30 photos per class on the Multimart Tomohon product dataset demonstrated convergence to 100% validation accuracy at epoch 4-5 under small, controlled dataset conditions; these results are condition-specific for Box 1 LED strip lighting and require full retraining each time a new product class is added. Real-time identification testing confirmed that the CNN dominates the identification of products with unobstructed packaging (94% confidence level), while OCR serves as a fallback for products with visible SKU codes or label-obscured packaging, with a system throughput of 66.7 FPS. The key contributions of this research are a two-box architecture that simultaneously eliminates sensor interference and enables automatic fraud detection through RFID-versus-infrared cross-verification, a conditional OCR/CNN pipeline capable of handling all product types regardless of SKU availability, and empirical evidence of system feasibility generalizable to small and medium retail warehouse contexts.