Night-Time Human Detection From UAV

The Unmanned Aerial Vehicle is an effective vehicle for rescue, search, and surveillance missions. A thermal camera improves the UAV system to operate these missions in the nighttime. Real-time human detection is an algorithm to increase performance and improve to be fully autonomous in rescue missions. Many studies have led to the integration of realtime human detection from thermal aerial images, but the task remains difficult from various human features from multi capture angle and UAV altitude. This paper proposes an experimental process for implementing real-time human detection from UAVs in the nighttime. We choose the YOLOv3 model for real-time human detection. Then, we create a custom thermal aerial human dataset that multi-capturing angle and altitude. The dataset is captured in the same condition of UAVs operation. We prepare and preprocess the dataset before sending it to the model training process. Finally, we evaluate a trained model for mean-Average Precision. The accuracy of prediction is evaluated with a test set and real- time detection performance. The results demonstrate that the model can detect a human in real-time with a thermal image from a UAV view and the accuracy of detection is mAP of 64.8% in the operating range of the UAV.