Feature-Based Detection Methods Of Small Targets In Sea Clutter

Maritime surveillance radar can cover a large pitch of sea surface by early warning, surveillance and detection, which has a wide application in the field of military and civilian areas. However, the non-stationary nature of sea clutter makes the conventional detection methods suffer a large performance decline and intolerable false alarm. It is important and difficult for the researchers both at home and abroad to detect the targets imbedding in sea clutter especially the small targets floating on the sea surface. Based on the analysis of real life sea clutter datasets, the main research of the dissertation is to exploit the detection methods for the floating small target imbedding in sea clutter. The research focuses on the fractal property of sea clutter and fractal-based detection method, tri-feather-based detection method, the detection method using the property of Doppler spectrum of sea clutter, block-whitened clutter suppression and target detection using time-frequency features. The major contributions of this dissertation are listed as follows,1. The multiscale fractal property and space-time varying fractal property of time series of sea clutter are analyzed. Based on the Extended Self-similar (ESS) model, the multiscale fractal property of sea clutter is discussed at three specific time scale ranges. The influence of noise on the fractal time series is discussed. Analysis of the IPIX radar datasets collected at different time shows that the Hurst exponent of sea clutter varies with sea state and the viewing geometry of the radar. By using the fractal property of sea clutter, two improved fractal-based detectors are given. The real life datasets show that compared with the Hurst parameter based detector, the multiscale Hurst parameter based and normalized Hurst parameter based detectors attain the better detection performances.2. The connection of the conventional target detection problem and one-class classification problem in anomaly detection is discussed, which provides a new solution in detecting the target in sea clutter. It is impossible to obtain the returns of all kinds of targets for the complexity and diversity of target on the sea surface. In one-class classification problem, we can treat the clutter-only returns as the normal observations and the returns with target as the anomaly observations. Three features that have the discrimination between the two patterns are extracted and the separability of the two patterns is analyzed in the 3D feature space. A fast convexhull learning algorithm is proposed to obtain the decision region from the feature vector of clutter-only observations. The tri-feature-based detector is proposed. Compared with the fractal-based detector, the proposed detector can attains a better performance.3. Based on the real life datasets, the Doppler spectrum of sea clutter is statistically modeled. Compared with the ground clutter, sea clutter has a relatively wide Doppler bandwidth. In Doppler domain, the Doppler bins are grouped into clutter-dominated bins, noise-dominated bins, and noise-clutter-mixed bins. Furthermore, the power variations at these bins are characterized by the K-distributions with different shape and scale parameters. In view of the Doppler spread when the target return is integrated within a long observation time, a double detection scheme is developed. The experimental results for the real life datasets show that the detector has a good performance when the SCR is high.4. The reference cells are not enough for the conventional whitening method when estimating the covariance matrix of sea clutter if the observation time is long because of the non-stationary nature of high-resolution sea clutter. Therefore, the clutter cannot be suppressed effectively and the detection performance is limited. To this end, block-whitened strategy is used to suppress the clutter and the detection under sea clutter background is transformed into the detection under white noise background. In this way, the influence of the cross-term on the target return is declined on the time-frequency plane (TF plane). Two detection method are proposed:(1) the method based on the TF-ridge-aided Hough transform (RAHT). Compared with the normal Hough transform the RAHT has the lower computational complexity and a good integration of target power as well. Thus, the method has the practical application value. (2) A joint-feature detection method via improved convexhull learning algorithm. Based on the prior knowledge of the distribution of the feature, the improved convexhull learning algorithm is proposed to determine the decision region. Aiming at the discrimination of the clutter-only return and the return with target on the TF plane, two approaches to extract the power of ridge and the total variation of ridge as the features to tell the target from sea clutter. Bhattacharyya distance is used to evaluate the separation of the two pattern of the two approaches on the TF plane. The two features obtained from the two approaches have the complementation in detecting the target imbedding in sea clutter and the real life datasets verify the efficiency of the two detectors.