Using computer vision to recognize composition of construction waste mixtures: A semantic segmentation approach

Timely and accurate recognition of construction waste (CW) composition can provide yardstick information for its subsequent management (e.g., segregation, determining proper disposal destination). Increasingly, smart technologies such as computer vision (CV), robotics, and artificial intelligence (AI) are deployed to automate waste composition recognition. Existing studies focus on individual waste objects in well-controlled environments, but do not consider the complexity of the real-life scenarios. This research takes the challenges of the mixture and clutter nature of CW as a departure point and attempts to automate CW composition recognition by using CV technologies. Firstly, meticulous data collection, cleansing, and annotation efforts are made to create a high-quality CW dataset comprising 5,366 images. Then, a state-of-the-art CV semantic segmentation technique, DeepLabv3+, is introduced to develop a CW segmentation model. Finally, several training hyperparameters are tested via orthogonal experiments to calibrate the model performance. The proposed approach achieved a mean Intersection over Union (mIoU) of 0.56 in segmenting nine types of materials/objects with a time performance of 0.51 s per image. The approach was found to be robust to variation of illumination and vehicle types. The study contributes to the important problem of material composition recognition, formalizing a deep learning-based semantic segmentation approach for CW composition recognition in complex environments. It paves the way for better CW management, particularly in engaging robotics, in the future. The trained models are hosted on GitHub, based on which researchers can further finetune for their specific applications.

Automated summary

The research automates construction waste composition recognition using CV technologies, with the development of a high-quality CW dataset, introduction of advanced CV semantic segmentation technique, and testing of training hyperparameters to improve model performance. The approach demonstrates high accuracy and robustness in experiments.