This repo is an implementation of our paper:
Abstract
A framework was developed using state-of-the-art deep learning methods for accurately counting materials in a low-light indoor construction jobsite environment. The Yolo v4 algorithm was improved to achieve a short-time convergent and accurate detection of the site operatives and materials. Then, DenseNet was deployed for classifying and recognising the objects. Finally, MC module based on morphology operations and the Hough transform was applied to automate counting materials within each stack.
Main steps of the framework are:
- Detection
- Classification
- Couting
- Prediction
In the following image, you can see the steps of the farmework with more details.
For object detection, we used Yolov4 code as a basic code and then improving loss function on it. For running Yolov4 on own dataset, we modified obj.name, obj.dat, yolov4.cfg, and src/. We put google colab code for this part on the Detection folder.
We finetuned DenseNet-161 in keras platform for our three classes of materials (0: left, 1: frontal, and 2: right). We put the finetuned denseNet for own data, implementing google colab on the Classificatin folder.
Site operatives counted by the results of the detection phase. However, the materials counted by MC module, which is based on the morphology operations, Hough Transform, and post processing algorithm.
Finally, our framework could predict the rate of waste, installed, and imported materials and also the number of operative sites in a scene.
In the following image, you can the results of the framework.
