Key points detection algorithm for noised data

This works introduces a new algorithm for feature detection in noised data, independently from the dimension of the given data. The algorithm is based on the detection and isolation of large features and its operability is demonstrated in this thesis through the development of two techniques, based on it. The first uses the algorithm for features detection on images, using image stitching as metrics for comparison with existing techniques. It demonstrates excellent performances on tests datasets registering a success rate almost three times higher than existing techniques while being fast and presenting a unique characteristic in the amount of points it detects for homography, largely inferior in number but superior in quality when compared to other techniques. The second technique demonstrate the performances achievable by the algorithm for feature detection on time series, it was developed in the framework of the SmartLINAC project at CERN. The technique showed excellent performance, detecting consistently all areas of anomalies, and labelling them correctly, where existing techniques showed large amount of false positive and false negative labelling entries due to the noise present in the data. The algorithm’s core concept is to ignore ambient noise in the data by a series of pre-processing techniques involving normalization, smoothing and thresholding, using noise’s statistical distribution’s attribute. Large areas are then isolated by blocks which’s characteristics can be used for comparison. The two techniques showed excellent performance in their range of application, proving the algorithm proposed in the thesis relevant and performant in its domain of application.