This paper explores the use of thermal imaging for search and rescue (SAR) missions using drones. The process involves object detection and multiple-target tracking. The You-Only-Look-Once (YOLO) model is used for object detection in thermal videos, while the Kalman filter is employed for track initialization, maintenance, and termination. A bounding box gating rule is proposed to enhance measurement-to-track association, and a statistically nearest neighbor association rule is combined with it. The proposed method is tested on three videos of hikers simulated as lost in the mountains, under challenging conditions such as low ambient light, close and occluded objects, and arbitrary drone movements. The results show robust tracking performance in terms of average total track life (TTL) and average track purity (TP), although the average mean track life (MTL) is reduced in harsh environments due to track breakage. The paper also discusses the limitations and future improvements, including the use of higher iterations of the YOLO model for better detection performance.This paper explores the use of thermal imaging for search and rescue (SAR) missions using drones. The process involves object detection and multiple-target tracking. The You-Only-Look-Once (YOLO) model is used for object detection in thermal videos, while the Kalman filter is employed for track initialization, maintenance, and termination. A bounding box gating rule is proposed to enhance measurement-to-track association, and a statistically nearest neighbor association rule is combined with it. The proposed method is tested on three videos of hikers simulated as lost in the mountains, under challenging conditions such as low ambient light, close and occluded objects, and arbitrary drone movements. The results show robust tracking performance in terms of average total track life (TTL) and average track purity (TP), although the average mean track life (MTL) is reduced in harsh environments due to track breakage. The paper also discusses the limitations and future improvements, including the use of higher iterations of the YOLO model for better detection performance.