Concordia University

http://www.concordia.ca/content/shared/en/news/encs/info-systems-eng/defences/2019/04/100/analysis-performance-hog-cnns-detecting-construction-equipment-personal-protective-equipment.html

notice

Master Thesis Defense - April 10, 2019: Analysis of the Performance of HOG and CNNs for Detecting Construction Equipment and Personal Protective Equipment

April 5, 2019

 

Seyedeh Forough Karandish

Wednesday, April 10, 2019 at 10:00 a.m.
Room EV002.184

You are invited to attend the following M.A.Sc. (Quality Systems Engineering) thesis examination.

Examining Committee

Dr. F. Naderkhani, Chair
Dr. A. Hammad, Supervisor
Dr. J. Y. Yu, CIISE Examiner
Dr. M. Nik-Bakht, External Examiner (BCEE)

 

Abstract

The construction industry remains one of the most dangerous working environments in terms of fatalities and accidents. High numbers of accidents and loss-time injuries, leads to a decrease in productivity in this industry. Therefore, new technologies are being developed to improve the safety of construction sites. Object detection on construction sites has a huge impact on the construction industry. Many researchers studied productivity, safety and project progress. However, few efforts have been made to improve the robustness of the related datasets for detection purposes. In the meantime, it is noticed that the lack of a custom dataset leads to low accuracy and also an increase in the cost and time of training dataset preparation.

In this research, we first investigated the generation of synthetic images using 3D models of construction equipment to use them as the datasets for training purposes, namely: excavators, loaders and trucks, and then sensitivity analysis is applied. We compared the performance of CNNs and other conventional methods for classifying construction equipment. In the second part, the detection of personal protective equipment for construction workers was studied. For this purpose, several object detection architectures from the TensorFlow object detection model zoo have been evaluated to find the best and most robust detection model. The dataset used in this study contains real images from construction sites. The performance evaluation of trained object detectors are measured in terms of mean average precision. The test results from this study showed that (1) synthetic images have a significant effect on the final detection results; and (2) comparing various object detection architectures, Faster-rcnn_resnet101 was the most suitable model in terms of accuracy of detection.

 




Back to top

© Concordia University