| Summary: | The unsupervised detection of maritime targets in grey scale video is a difficult problem in maritime video surveillance. Most approaches assume that the camera is static and employ pixel-wise background modelling techniques for foreground detection; other methods rely on colour or thermal information to detect targets. These methods fail in real-world situations when the static camera assumption is violated, and colour or thermal data is unavailable. In defence and security applications, prior information and training samples of targets may be unavailable for training a classifier; the learning of a one class classifier for the background may be impossible as well. Thus, an unsupervised online approach that attempts to learn from the scene data is highly desirable. In this thesis, the characteristics of the maritime scene and the ocean texture are exploited for foreground detection. Two fast and effective methods are investigated for target detection. Firstly, online regionbased background texture models are explored for describing the appearance of the ocean. This approach avoids the need for frame registration because the model is built spatially rather than temporally. The texture appearance of the ocean is described using Local Binary Pattern (LBP) descriptors. Two models are proposed: one model is a Gaussian Mixture (GMM) and the other, referred to as a Sparse Texture Model (STM), is a set of histogram texture distributions. The foreground detections are optimized using a Graph Cut (GC) that enforces spatial coherence. Secondly, feature tracking is investigated as a means of detecting stable features in an image frame that typically correspond to maritime targets; unstable features are background regions. This approach is a Track-Before-Detect (TBD) concept and it is implemented using a hierarchical scheme for motion estimation, and matching of Scale- Invariant Feature Transform (SIFT) appearance features. The experimental results show that these approaches are feasible for foreground detection in maritime video when the camera is either static or moving. Receiver Operating Characteristic (ROC) curves were generated for five test sequences and the Area Under the ROC Curve (AUC) was analyzed for the performance of the proposed methods. The texture models, without GC optimization, achieved an AUC of 0.85 or greater on four out of the five test videos. At 50% True Positive Rate (TPR), these four test scenarios had a False Positive Rate (FPR) of less than 2%. With the GC optimization, an AUC of greater than 0.8 was achieved for all the test cases and the FPR was reduced in all cases when compared to the results without the GC. In comparison to the state of the art in background modelling for maritime scenes, our texture model methods achieved the best performance or comparable performance. The two texture models executed at a reasonable processing frame rate. The experimental results for TBD show that one may detect target features using a simple track score based on the track length. At 50% TPR a FPR of less than 4% is achieved for four out of the five test scenarios. These results are very promising for maritime target detection.
|
|---|
