Disaster Victims Detection: CNN as Feature Extraction for Ensemble Learning by Utilizing Random Tree

Disaster is something that we, human, cannot predict, and can occur anytime without us knowing. Disaster will cause many losses, in material side, or in human victims. Human that buried under the debris is in a critical state that needs to be rescued as soon as possible. The problem arises when detecting the victims, because the limit of human ability to detect victims among the debris, that is often inaccurate, and takes too long, that often exceeded the golden hours. Developing an accurate and efficient object detection is a must. We will try to classify images based on the 4 class that is available which is ‘utuh’, ‘kepala’, ‘tangan’, and ‘kaki’ which represents the real-time condition. This research will use various CNN models like ResNet-50, Inception-V3, EfficientNet, DenseNet-121, and VGG19 to extract features from image. We will also try to discover and solve the problem with the trade-off between accuracy and FPS, that is important in terms of real-time object detection. The result has shown the outstanding performance of VGG19, which outperforms all models. VGG19 achieve 150.16 for FPS which is a high score that is almost double the score of the rest models FPS. The accuracy of VGG19 is the same as EfficientNet which is 0.9840.