Object detection and classification using matched filtering and higher-order statistics

Summary form only given. If a known object is corrupted by additive white Gaussian noise, then the matched filter maximizes the output signal-to-noise ratio. The main drawback of the matched filter in two dimensions is its sensitivity to object shifts, rotation, and scaling, especially in the presence of additive colored Gaussian noise of unknown covariance. These problems have been overcome by using higher-order statistics (HOS). The zero lag of the triple correlation of the matched filter output has been computed and compared with zero. Since the triple correlation of a Gaussian process is zero, it has been shown that this statistic will peak if the object is present. A detection algorithm that exploits all the output samples of a single matched filter has been developed. Rotation and scaling invariance have been incorporated by transforming the Cartesian coordinates of the image and the templates into log-polar coordinates.