Optical Load Detection: Load Weighing for Construction Machines using Stereo Vision and Convolutional Neural Networks

Abstract

Accurate excavation monitoring is important for the handling of materials within the construction industry. Modern construction machines provide built-in systems for weighing handled goods. In this thesis, an alternative optical weighing system is developed and implemented for an excavator and a wheel loader. The optical system detects and provides the volume and weight of the handled material through fill-factor estimation. The methodology is based on depth data and images captured by a stereo camera, mounted on the machines. By using a region-based convolutional neural network (CNN), localization of material and fill-factor estimation are managed jointly. Material classification is also proved to be possible using gathered images and a simple CNN. By combining the fill-factor and information about the material, weight is obtained. Evaluations reveal that the system measures fill-factor to mean absolute percentage errors (MAPE), relative to the maximum capacity of the excavator and the wheel loader, of 3.3 % and 3.0 % respectively.