Object Detection and Classification for Mobile Platforms Using 3D Acoustic Imaging

This thesis addresses the problem of obstacle detection and classification for mobile platforms such as robots using acoustic imaging. To obtain an acoustic image of a scene, a spatially broad signal is transmitted and the ob ject’s reflections are received by a 2D array of acoustic receivers. The resulting data is processed using adaptive beamforming and can be translated into a three-dimensional image. Since the targeted platforms operate in man-made environments, we first develop design principles derived from physical constraints such as the slow speed of sound in air. Furthermore, we present a calibration method which is specifically well suited for acoustic arrays and parametrically corrects for position errors of the sensors. Other methods are either limited, e.g., by the need of a high number of calibration sources, or can correct such errors non-parametrically and therefore sometimes insufficiently. The increasing cost pressure for domestic robots demands to operate using cheap hardware, which favors the use of acoustic imaging. Such cost constraints require to use highly sparse 2D array designs which still allow to resolve ob jects clearly and result in acoustic images with a distinct discrimination between ob ject echoes and background. Thus, we develop methods to design non-uniform, sparse arrays which possess reasonable spatial resolution together with good noise suppression. The presented methods apply minimum-redundancy theory in the two-dimensional case and extend it in order to control the redundancy. We also address the problem of human detection and develop feature sets for a corresponding binary classifier. As a result, humans can be discriminated from other ob jects using only a three-dimensional feature space and simple classifiers such as Linear Discriminant Analysis or Quadratic Discriminant Analysis with an accuracy of almost 97 percent. We also present geometrical and statistical features which allow the classification of humans with respect to their pose, meaning that we can distinguish whether a person is walking or standing with a classification accuracy of more than 87 percent. All developed methods are applied not only to simulation data, but also to real data measurements. The data was obtained using several prototypes of real acoustic array systems in indoor and outdoor environments