Vision-based method to identify materials transported by dump trucks

Abstract

Construction and demolition waste (C&DW) materials are highly valuable as, in most cases, they can be reused. In other cases, such material can be highly pollutant. Therefore, their traceability becomes very important in the frame of sustainable development. Computer vision-based solutions have the potential to recognize different forms of construction waste material. This study provides a vision-based framework to identify C&DW transported by dump trucks before entering landfills. The proposed framework is composed of two main models: a semantic segmentation model to isolate the content of the dump truck, and a classification model to identify the C&DW in the image patches created by the superimposition of the segmentation mask and the input image. The results of the classification model are used to assess the homogeneity of the content and the main type of waste material contained in the dump truck. The semantic segmentation model achieves an IoU of 0.945 and an average inference time of 0.388s on a CPU, and the classification model achieves a F1-score of 0.965. To validate the proposed framework, various data collection attempts have been conducted in real landfills. The validation tests achieved an average accuracy of 99.0% in identifying the main material in the dump truck and an average accuracy of 94.8% in evaluating its homogeneity. The developed framework performed its inference in 0.899 s on a CPU.