In this study, we implement a robust algorithm to locate victims in disaster scenarios caused by collapsed buildings using computer vision techniques. We propose a hybrid algorithm that combines transfer learning, combining the YOLO object detection model and the MediaPipe pose estimation model. This work was developed in stages, including data collection and preparation, model training, model fine-tuning, robustness using pose estimation, as well as validation and testing. As a result, we achieved an average overall accuracy of 82.7% in victim detection with a pose classification capacity of 85%, confirming its applicability in real-world environments. We consider adverse conditions with poor lighting, vision of only body parts, and unconventional poses of victims. The system demonstrates robust performance, with an efficiency of 79% in detecting partially visible victims and 76% in scenarios with dense debris. This work contributes to the development of autonomous technologies applied to rescue missions, demonstrating the feasibility of the hybrid approach.