Unmanned Aircraft Systems (UAS) represent a significant potential for growth in the aerospace industry. Their use in military operations has increased exponentially in the last decade alone, requiring a corresponding increase in training airspace in the United States. In addition to military usage, UAS have the potential to fulfill a myriad of roles for both the public and private sectors. However, the use of UAS has been limited in the National Airspace System (NAS) to military and public applications and only under fairly restrictive Certificates of Authorization or Waiver (COA). The only way to truly realize the potential of UAS is to fully integrate them into the NAS. The desire to integrate UAS was recently codified into law with the 2012 FAA Modernization Act, mandating integration by specific, fairly short timelines. There are several challenges currently preventing the full integration of UAS that range from technological to procedural areas. However, the one common theme in all of these challenges is Safety. Across the literature on this topic there is no consensus on how safe UAS need to be to achieve integration, whether UAS can currently meet specified safety targets, and if not, what is the best way to achieve the safety goals. The purpose of this effort was to demonstrate a comprehensive framework for analyzing UAS integration efforts using a Target Level of Safety (TLS) approach. Using reliability tools, aircraft encounter models, and data from a wide variety of sources ranging from manned aircraft safety, explosives, falling debris and earthquake damage, the primary outcome of the effort was a better understanding of the risk to second and third party persons as a result of UAS operations in the NAS. This framework and associated models are validated using reliability and casualty data from manned aircraft operations. The framework is then applied to several relevant and specific cases to demonstrate the impact of policy decisions on UAS reliability and allowed operational areas. The supporting research and analysis can serve as a baseline for future integration analysis and decision-making efforts, and was designed to allow stakeholders and decision makers in this field to assess UAS safety, and set minimum system reliability requirements and mitigation system effectiveness standards.

A forced landing for an Unmanned Aerial Vehicle (UAV) is required if there is an emergency on board that requires the aircraft to land immediately. Piloted aircraft in the same scenario have a human on board that is able to engage in the complex decision making process involved in the choice of a suitable landing location. If UAVs are to ever fly routinely in civilian airspace, then it is argued that the problem of finding a safe landing location for a forced landing is an important unresolved problem that must be addressed. This thesis presents the results of an investigation into the feasibility of using machine vision techniques to locate candidate landing sites for an autonomous UAV forced landing. The approach taken involves the segmentation of the image into areas that are large enough and free of obstacles; classification of the surface types of these areas; incorporating slope information from readily available digital terrain databases; and finally fusing these maps together using a high level set of simple linguistic fuzzy rules to create a final candidate landing site map. All techniques were evaluated on actual flight data collected from a Cessna 172 flying in South East Queensland. It was shown that the use of existing segmentation approaches from the literature did not provide the outputs required for this problem in the airborne images encountered in the gathered dataset. A simple method was then developed and tested that provided suitably sized landing areas that were free of obstacles and large enough to land. The advantage of this novel approach was that these areas could be extracted from the image directly without solving the difficult task of segmenting the entire image into the individual homogenous objects. A number of neural network classification approaches were tested with the surface types of candidate landing site regions extracted from the aerial images. A number of novel techniques were developed through experimentation with the classifiers that greatly improved upon the classification accuracy of the standard approaches considered. These novel techniques included: automatic generation of suitable output subclasses based on generic output classes of the classifier; an optimisation process for generating the best set of input features for the classifier based on an automated analysis of the feature space; the use of a multi-stage classification approach; and the generation of confidence measures based on the outputs of the neural network classifiers. The final classification result of the system performs significantly better than a human test pilot's classification interpretation of the dataset samples. In summary, the algorithms were able to locate candidate landing site areas that were free of obstacles 92.3 ±2.6% (99% confidence in the result) of the time, with free obstacle candidate landing site areas that were large enough to land in missed only 5.3 ±2.2% (99% confidence in the result) of the time. The neural network classification networks developed were able to classify the surface type of the candidate landing site areas to an accuracy of 93.9 ±3.7% (99% confidence in the result) for areas labelled as Very Certain. The overall surface type classification accuracy for the system (includes all candidate landing sites) was 91.95 ±4.2% (99% confidence in the result). These results were considered to be an excellent result as a human test pilot subject was only able to classify the same data set to an accuracy of 77.24 %. The thesis concludes that the techniques developed showed considerable promise and could be used immediately to enhance the safety of UAV operations. Recommendations include the testing of algorithms over a wider range of datasets and improvements to the surface type classification approach that incorporates contextual information in the image to further improve the classification accuracy.