Sonar image interpretation for sub-sea operations
Daniell, Oliver James
MetadataShow full item record
Mine Counter-Measure (MCM) missions are conducted to neutralise underwater explosives. Automatic Target Recognition (ATR) assists operators by increasing the speed and accuracy of data review. ATR embedded on vehicles enables adaptive missions which increase the speed of data acquisition. This thesis addresses three challenges; the speed of data processing, robustness of ATR to environmental conditions and the large quantities of data required to train an algorithm. The main contribution of this thesis is a novel ATR algorithm. The algorithm uses features derived from the projection of 3D boxes to produce a set of 2D templates. The template responses are independent of grazing angle, range and target orientation. Integer skewed integral images, are derived to accelerate the calculation of the template responses. The algorithm is compared to the Haar cascade algorithm. For a single model of sonar and cylindrical targets the algorithm reduces the Probability of False Alarm (PFA) by 80% at a Probability of Detection (PD) of 85%. The algorithm is trained on target data from another model of sonar. The PD is only 6% lower even though no representative target data was used for training. The second major contribution is an adaptive ATR algorithm that uses local sea-floor characteristics to address the problem of ATR robustness with respect to the local environment. A dual-tree wavelet decomposition of the sea-floor and an Markov Random Field (MRF) based graph-cut algorithm is used to segment the terrain. A Neural Network (NN) is then trained to filter ATR results based on the local sea-floor context. It is shown, for the Haar Cascade algorithm, that the PFA can be reduced by 70% at a PD of 85%. Speed of data processing is addressed using novel pre-processing techniques. The standard three class MRF, for sonar image segmentation, is formulated using graph-cuts. Consequently, a 1.2 million pixel image is segmented in 1.2 seconds. Additionally, local estimation of class models is introduced to remove range dependent segmentation quality. Finally, an A* graph search is developed to remove the surface return, a line of saturated pixels often detected as false alarms by ATR. The A* search identifies the surface return in 199 of 220 images tested with a runtime of 2.1 seconds. The algorithm is robust to the presence of ripples and rocks.