A Deep Learning and Vision-Based Solution for Material Volume Estimation Considering Devices’ Applications

Abstract

The estimation of material volume in a construction vehicle’s bucket is a crucial prerequisite for automation, as well as for productivity assessment and efficient material transport. Although some studies have been conducted in this field, the accuracy and speed of inference have been suboptimal, and specific implementation strategies have not been proposed. To address these issues, this paper proposes a new approach. The proposed approach has three main components. First, a novel image preprocessing framework based on three-dimensional (3D) grayscale terrain is presented. Second, a semantic mask-level data set is constructed to facilitate future research in this area. Third, a combined neural network and probabilistic approach is proposed to estimate the material volume, with speed and accuracy as metrics. Transfer learning is introduced to improve training efficiency and accuracy. The proposed material volume estimation method is implemented on three different devices, addressing the problem from the development phase to the application phase. The advantages and disadvantages of each device are discussed in depth. The results demonstrate that the proposed approach achieves an impressive average accuracy of 98.20% on all three devices, with real-time or semi–real-time volume estimation feasible on each. In summary, this paper proposes a new approach to estimate the material volume in a construction vehicle’s bucket, addressing issues of accuracy and speed of inference and providing specific implementation strategies. The results demonstrate the effectiveness of the proposed approach, which has potential applications in automation and productivity assessment in the construction industry.