Human detection for search and rescue operations using embedded artificial intelligence

Abstract

Unmanned aerial vehicles (drones) have been increasingly used in search and rescue operations as a tool to detect humans in an area of disaster where the rescue team is unable to reach them. Human detection is the most important task in a rescue plan. Currently, deep learning and Internet of Things (IoT) technologies are used to automatically detect humans from footage taken from drones, however, the hardware used in such methods consumes high power, requires high processing capability, long computational time, and a constant internet connection which are not effective to be deployed in all scenarios. This project aims to utilize transfer learning to build a human detection model with mean average precision (mAP@0.5) above 90% and compare deep learning models in aspects of the size of the model, computational requirement, and mAP. Furthermore, to compress the final model to be deployed to an edge device for the propose of using edge computing where the computation requirement for the deep learning model is all made on-chip. The development of this project is based on multi-datasets using the TensorFlow v2 framework and virtual machine Google Colaboratory. The dataset used in this project is extracted from two datasets named SARD and SeaDroneSee both are aerial images of humans, and the labeling of the dataset is made using the Roboflow platform. The models used are pre-trained single shot detector models namely MobileNet v2 and EfficientDet-D1, the last shows a better accuracy of 97.3% mAP@0.5 however MobileNet v2 consumes much less GPU for training at around 4.6 GB while maintaining relatively high accuracy of 95.5% mAP@0.5. Lastly, the trained MobileNet v2 model is quantized to 6.4 MB. At the end of this project, a deep learning model for human detection for search and rescue operations is compressed and ready to be deployed to an embedded artificial intelligence device.