MSc Thesis Defense by Mads Thoudahl
Machine Learning for Naval Search and Rescue - by Distance, Intensity and Behavior Features.
Present thesis investigates usage of machine learning methods for pixel classification in a naval aerial search and rescue context. The footage used as data source is recorded using a long wave infra red sensor, registering thermal intensities in high definition resolution.
A training data set of 2000 samples equally distributed in 5 categories is manually categorized. A survey is used to select an artificial neural network model, when investigating performance in pixel classification. A manual pixel classification of 3 key cases is performed and used as ground truth in performance analysis.
A novel feature set denoted Distance Intensity Behavior (DIB) is proposed, and implemented. The feature set consist of a distance layer, in which distance to the object represented by each individual pixel is estimated. The intensity layer consist of the raw intensities at the current video frame. The behavior layer is an estimate of the spatial locality differentiated with regard to time dimension. Training data is reorganized in image patches, and distance and behavior descriptors calculated. The number of samples are increased by collecting a few new samples, and by mirroring the existing samples, doubling their numbers. Background categories are merged after prediction in either case.
Two individual neural networks are trained, one using Raw intensity features, the other using the DIB feature set. Model performance statistics are calculated and evaluated using the key cases. The results show a substantial increase in performance in long and medium range cases. The performance in the close range case is unchanged at a significantly lower level. With an error rate of only 33 false positives, it is shown that it is possible to identify 2 of 4 person in water pixels in more than 1.1 million pixels. Identifying a vessel is also shown to be viable with a great precision as 393 of 423 vessel pixels are positively identified with only 15 false positives in the same 1.1 million pixels.
Supervisor: Jon Sporring, DIKU
Exterminal examiner: Jens Damgaard Andersen, DIKU