Development of Landslide Victim Detection System using Thermal Imaging and Histogram of Oriented Gradients on E-PUCK2 Robot

Thermal imaging currently have been widely used in many aspects; one of them is in search and rescue application. The system utilizes thermal imaging as a detection system to differentiate between living human and post-landslide environment. This paper proposes the use of a histogram of oriented gradient (HOG) features from the images taken by the thermal imaging camera to be used to classify a landslide victim (target) from its surrounding environment. The model is built by three different algorithms, i.e., support vector machine (SVM), K-nearest neighbor, and random forest. Those algorithms were used to classify the target from its surrounding based on the image taken by FLIR Camera. Based on the simulation results, the model built with 5×5 pixels of HOG feature and SVM with linear kernel achieved 81.82% accuracy and has the lowest computation time among other algorithms, i.e., 0.526 seconds per identified image. This best model is then implemented and tested on an EPUCK2 swarm robot as a search robot in the designed landslide search and rescue scenario. Experimental testing results show that the developed system can successfully classify seven targets out of 12 different experiments with 72.43% accuracy.