Assisting Avalanche Search and Rescue Operations with UAV Imagery
A paper by Bejiga et al. (2017) introduces and investigates the success of an automated system that utilises images captured by an Unmanned Aerial Vehicles (UAV) to detect avalanche victims.
According to a report by Brugger & Falk (2002), around 60% of avalanche victims have at least some part of their body visible. The present study aims to to detect such victims and therefore help speed up the search part of a Search And Rescue operation by utilising the following automated steps:
- pre-process the images to select regions of interest;
- employ a Convolutional Neural Network technique to identify features within the images;
- train a Support Vector Machine (SVM) to classify what type of object the features are;
- post-process the results, using Hidden Markov Models to compare whether objects appear in successive images as would be expected.
There are currently two possibilities as to how this system would be deployed: either the images are sent to a ground station in near-real-time and processing is achieved there, which is slower but more powerful; or the processing takes place onboard the UAV, and only areas of interest are sent to the station.
And in a little more detail…
Pre-processing: Image segmentation methods split images into areas of snow and non-snow, based on a threshold value that determines their colour.
Convolutional Neural Network: A feed forward neural network - neurons accept input from local neurons in the previous layer, and use weighted operations based around the width, height and depth of the neuron. The convolutional layer is trained to identify features, the pooling layer mitigates against overfitting, and the fully connected layer introduces the weighting of all other neurons from the previous layer.
Support Vector Machine: Identified features from the CNN are classified, using margin maximization (between classes) and error minimization (penalising incorrect classification). The aim is to separate people and skis from trees, shadows and rocks.
Post-processing: Use of HMM to compare successive images. These techniques are used in applications such as speech recognition, molecular biology, and bioinformatics.
The use of a pre-processing step significantly decreases the processing time.
Increasing the resolution of the input data improves the detection success of the system, but also leads to a larger number of false alarms and slows down the processing time.
Increasing the height from which the images are captured reduces the prediction performance.
Investigate a pre-processing step that is more robust at lower resolutions.
Investigate the effects of illumination and wind on the images.
Bejiga, M.B., Zeggada, A., Nouffidj, A. & Melgani, F. (2017) A Convolutional Neural Network Approach for Assisting Avalanche Search and Rescue Operations with UAV Imagery. Remote Sensing, 9, 2, 100. Open Access Article here.
Brugger, H. & Falk, M. (2002) Analysis of Avalanche Safety Equipment for Backcountry Skiers. Available online here. Accessed on 26 February 2017.
