A Deep Learning Framework for Object Detection and Tracking: CNN-Based Parcel Tracking on Conveyor Belts

Object Detection and Tracking (ODT) represents a critical area within computer vision, with broad applications across various industries, particularly in the automation of logistics operations. However, significant challenges persist, including the logistics-specific datasets and the inefficiency of data annotation, which is crucial for training robust artificial intelligence models for parcel detection and tracking. These challenges are exacerbated when human experts are required to manually assign unique identifiers to objects across frames, hindering the scalability of AI model training. In this research, we present a novel deep learning-based ODT framework, leveraging Convolutional Neural Networks (CNN) integrated with evaluation metrics such as Euclidean Distance and Intersection over Union (IoU) to facilitate robust object detection and tracking. To address the limitations of manual annotation, we develop an automated framework for assigning unique identifiers to bounding boxes across sequential frames. The framework is based on three logistics-related datasets, comprising time-series images of packages on moving conveyor belts. The first two datasets, containing 100 and 1,021 images, are manually annotated, while the third dataset of 5,046 images is automatically annotated using the proposed tool. We also utilize the Hungarian algorithm to optimize identity assignment across frames. The results demonstrate that the framework achieves an accuracy of 99.09% in object detection and tracking with a precision of 99.25% and a recall of 99.10% for the manually annotated dataset. The automated annotation illustrates an accuracy of 98.70%. These findings indicate the robustness and scalability of the proposed framework in automating parcel tracking for logistics operations.