CNN-based Human Detection for UAVs in Search and Rescue

TL;DR

This paper presents a UAV-based search and rescue system using a CNN (SSD model) on a Raspberry Pi, demonstrating effective victim detection from above with stable flight and real-time processing capabilities.

Contribution

It introduces a practical implementation of SSD-based human detection on a DIY UAV platform with real-world testing and noise handling techniques.

Findings

Stable flight achieved with noise filtering

SSD model runs at 3 fps on Raspberry Pi B

Effective detection within 1 to 20 meters

Abstract

The use of Unmanned Aerial Vehicles (UAVs) as a substitute for ordinary vehicles in applications of search and rescue is being studied all over the world due to its flexible mobility and less obstruction, including two main tasks: search and rescue. This paper proposes an approach for the first task of searching and detecting victims using a type of convolutional neural network technique, the Single Shot Detector (SSD) model, with the Quadcopter hardware platform, a type of UAVs. The model used in the research is a pre-trained model and is applied to test on a Raspberry Pi model B, which is attached on a Quadcopter, while a single camera is equipped at the bottom of the Quadcopter to look from above for search and detection. The Quadcopter in this research is a DIY hardware model that uses accelerometer and gyroscope sensors and ultrasonic sensor as the essential components for balancing control, however, these sensors are susceptible to noise caused by the driving forces on the model, such as the vibration of the motors, therefore, the issues about the PID controller, noise processing for the sensors are also mentioned in the paper. Experimental results proved that the Quadcopter is able to stably flight and the SSD model works well on the Raspberry Pi model B with a processing speed of 3 fps and produces the best detection results at the distance of 1 to 20 meters to objects.