Machine Learning Driven Search and Rescue.

点击翻译按钮获取中文摘要

Search and Rescue (SAR) of vulnerable missing persons is unfortunately a common task for the Police and otheremergency services. Organisations like the Centre for Search and Rescue (CSR)[1] carry out research into thetypical behaviour patterns for classes of missing person (young child, elderly person with dementia etc.) andprovide specialist training to Police forces in all areas related to SAR. With the rapid development and expansion of the drone sector over the last decade, Unmanned Aerial Vehicles (UAVs) have become cheaper and more accessible than ever before. Officers from the Air Support Unit of Police Scotland are one of the first emergency services in the UK to take advantage of this technology by using infrared and visible light cameras on commercially available quadrotors as an aid to the current single manned helicopter. Current plans are to use single UAV platforms to search a pre-defined area, to detect possible targets and direct the search team, although swarms of vehicles are also being considered for future use. During some proof-of-concept research, numerical optimisation methods such as genetic algorithms and particle swarm optimisation were used to determine optimal paths. These paths were shown to yield faster search times when compared to the classic parallel swaths (lawnmower pattern) and even compared to trained Police Scotland Air Support Unit operators. This research project intends to extend and automate the extremely promising results already obtained using manual creation of probability maps through the use of strong AI and machine learning. In particular, our research hypothesis is that a combination of fuzzy logic and deep learning could be used to combine the psychological descriptors of missing person classes (which are natural-language based) with the image classification capabilities of deep learning neural networks. The primary objective of using AI/machine learning would be to create new algorithms for automatic probability map creation, but we will also explore the benefits of using these techniques to replace, augment or reduce computation time of the path optimisation processes. The effectiveness of single and multiple UAVs will be explored, as will the constraints imposed by holonomic and non-holonomic platforms.