Implementasi Algoritma SSD MobileDets untuk Mendeteksi Aktivitas Operator Memasuki Area Hazard di Sekitar Mesin

Abstract

Abstract— Safety laser scanners are the main sensors proposed to protect work accidents in the industrial sector. However, they are expensive and limited to a 2D scanning area. Therefore, in this final project, research is carried out related to low-cost sensors using monocular cameras and Edge TPU as an alternative security laser scanner. The system can inform safety warnings using green, yellow, and red indicator lights, and use warning sounds using a buzzer in dangerous conditions. I captured the safe area using a monocular camera and processed the MobileDets SSD deep learning object detection image. To be able to detect objects, the person data that has been collected will be carried out a model training process, the model training process is carried out in order to produce a model that is ready for use. During the training process, the model obtained an mAP value of 0.217% on the validation data, and obtained an mAP value of 0.274% on the test data. Next, the object detector was run on Edge TPU, a USB device to accelerate the deep learning process. The resulting object detector model can detect objects in the form of people with an average accuracy of 73,81%. The object detector will recognize people and mark them with a bounding box. From the resulting bounding box, the object coordinates are then calculated to analyze people entering the safe area. When a person is detected entering a safe area, the system will provide a warning with the resulting indicator light. In addition, this system can generate security signals using USB Input/Output that can be integrated with machine or robot controllers