A Systematic Review of Deep Learning and Computer Vision Methods for Accurate Object Volume Measurement

Abstract

Precise and efficient object volume measurement is crucial across diverse industrial domains, including logistics, manufacturing, and agriculture, where traditional methods are often labor-intensive and error-prone. DL and CV technologies offer a compelling alternative, providing greater accuracy, speed, and safety through non-contact, real-time, and adaptive solutions. To address existing knowledge gaps, this SLR is believed to be the first to integrate and analyze the three critical dimensions of DL holistically- and CV-based volume measurement: core methodologies, practical applications across various industries, and outstanding challenges, thereby providing a unified and comprehensive understanding of the state of the art that previous, more fragmented reviews have failed to deliver. Regarding methodology, dominant DL techniques include CNNs, Mask R-CNN, and U-Net for segmentation; GANs for 3D model generation; and PointNet/voxel networks for 3D data processing, with sensor integration impacting model architecture. Applications span agriculture, logistics, manufacturing, and construction, demonstrating high accuracy with error rates as low as 0.75% and MAPE typically ranging from 3.2% to 5%. Challenges involve occlusions, diverse environmental conditions, data scarcity, and computational costs. Prioritized research directions include lightweight models, multi-task learning, improved generalization, greater robustness, and explainable AI. Overall, this SLR comprehensively synthesizes current DL and CV methodologies, their practical applications, and future research directions in object volume measurement.