Object Detection for Safety Attire Using Faster R-CNN

• Main Author: Amir Ikram Bin Hashim
• Language: English
• Published: 2023
• Subjects: Object Detection; Faster R-CNN; Safety Attire

Object detection and recognition is a computer vision technique in which to identify object of certain class such as human, animals, cars, tress and many more with a goal to teach and train the computer to be able to classify the detected object. Various advancement has been done in the field of object detection and recognition throughout the year. Studies done by the Department Occupational Safety and Health Malaysia showed that the number of accident that occurred in construction and manufacturing sector are not that satisfactory due to being the highest amount of cases compared to other sector and still can be reduce further. The main objective for this thesis is to create a working prototype using the object detection method with the ability to detect those who are not wearing proper safety attire while working in a hazardous area using Faster R-CNN object detection model. For an object detection and recognition to work, dataset should first be obtain. Dataset can be easily obtain from the internet or create using a camera. The dataset acquired will be used to train and validate the object detection model. The training for the model should stop after achieving a loss value of 0.1 and validate it using a webcam to see the performance of the model. Retraining of the model should be done if the accuracy rate are low or inaccurate detection (false detection). Various test have been conducted such as varying the amount of class and dataset in each class, varying the time taken to train the object detection model and comparing the performance of different object detection architecture such as Single Shot Multi-Box Detector (SSD). In addition, LabelIMG software was used to create the safety dataset. Anaconda software and Microsoft Visual Studio Code software was use to run, test and troubleshoot the object detection sources code. All the test was done using personal computer equipped with Nvidia GTX 1080 8GB graphic card and Logitech C310HD webcam. Comparison for the object detection architecture are done and it can be concluded that Faster R-CNN was chosen for this project as it can detect the safety attire correctly with high accuracy rate with less amount of training time compared to SSD. Further research and improvement are needed so that the object detection model proposed can be more versatile if placed at any location and be able to detect other object such as facial or even biometrics.