Camera-based Object Detection and Identification using YOLO Method for Indonesian Search And Rescue Robot Competition

Abstract

Indonesian Search And Rescue Robot Contest is a competition that competes for firefighting and rescue missions by a legged robot in a burning building simulation. The robot must be able to extinguish a candle flame and rescue the victim to a safe zone area as fast as possible. Many team members were already capable of completing the firefighting mission but still had trouble in the victim rescue mission. To detect victims, mostly used cameras and object detection algorithms like color detection on HSV color space. This algorithm only detects the color without knowing what object it is. There is a possibility that the algorithm detects incorrectly if detection is carried out in places with different light conditions or if there are other objects that have a similar color. In this paper, we present the implementation of YOLOv4-tiny on a Search And Rescue robot with limited hardware specifications and real-time case conditions. The YOLO model created has an mAP value of 98.25% from the training with 458 training data and 98 validation data. The model can detect 178 of 195 objects in validation data with a percentage of IoU is 79.23%. When embedded in Search And Rescue Robot, the robot can detect objects in real-time and get an average of 17.8 FPS. The YOLO model can detect objects in variations of light conditions, with the max distance of the victim and candle object around 1.5 m. With this high accuracy and speed, this algorithm is suitable for the Indonesian Search And Rescue Robot competition.