Design of a Tailored Edge AI System for Automated Box Counting in Real-Time Palletization

The automation of inventory tracking processes is a critical challenge in modern warehousing, driven by the growing demand for efficient, accurate, and adaptable logistics operations. Manual processes, still prevalent in many facilities, are time-consuming, error-prone, and ill-suited for the dynamic demands of contemporary supply chains. This thesis introduces a tailored system for automating box counting during palletization, integrating advanced computer vision techniques with low-cost hardware, including a Raspberry Pi 5 and a Time-of-Flight (ToF) sensor. The proposed approach addresses critical challenges such as irregular stacking patterns, diverse box dimensions and occlusions. By leveraging deep learning models optimized for edge devices, the system achieves real-time processing capabilities without the reliance on centralized computational infrastructure. The system employs a multi-stage pipeline that combines state-of-the-art object detection algorithms, depth data analysis, and real-time edge computing. This architecture ensures precise 3D box localization and dimension extraction, resulting in an accurate counting. Extensive optimization ensures the system operates efficiently on resource-constrained hardware, making it practical for real-world applications. Experimental validations in controlled and real-world environments show an average counting error of less than 2 boxes per 100. This research contributes to the advancement of warehouse automation by providing a scalable and efficient solution for industrial palletization. Its deployment can significantly reduce manual labor, streamline operations, improves inventory tracking precision, and set a foundation for further innovations in automated logistics.