A Novel Approach for Disaster Victim Detection Under Debris Environments Using Decision Tree Algorithms with Deep Learning Features

In order to maximise survival prospects, search and rescue (SAR) efforts in collapsed buildings are particularly difficult since they need to quickly identify victims within the first 48 hours. Automation is required due to the intricacy of unstructured settings and the significant risk associated with manual searches. In order to improve Human Victim Detection (HVD) in disaster-affected areas, this study suggests a Deep Learning method based on Transfer Learning. To train a deep learning model, a custom dataset was created with three class labels: hand, leg, and body. ResNet-50 was used to do feature extraction, utilising its prior training experience to increase accuracy with sparse data. In order to reduce duplicate information and increase computing performance, J48 decision tree pruning was used for feature selection in order to optimise classification. To identify the top-performing model, a number of machine learning classifiers were tested, including Random Forest, XGBoost, SVM, MLP, and Naïve Bayes. With an astounding accuracy of 99% and a calculation time of only 0.01 seconds, the XGBoost method performed better than any other algorithm, making it ideal for real-time applications.