Improved stochastic detection algorithms with applications in radar and Ladar based Robotic Mapping

Abstract

In this work a set of robust tools for noise reduction and feature detection in radar and ladar environment data for outdoor robotic applications have been developed. One of these tools is a new noise reduction technique for radar data which combines the well known spectral noise subtraction with the standard constant false alarm rate (CFAR) detectors and binary integration. The resulting method exhibited a relatively low time complexity compared to other state-of-the art noise reduction techniques, the Wiener filter and Wavelet denoising, while retaining a higher Signal-to-Noise ratio. The method was tested with real data from a local park captured with a scanning radar mounted on a robot platform and with SAR data available from NASA/JPL UAVSAR missions. The second component of this framework is the extension of the standard CFAR detectors used in radar data to be used with 3D ladar point cloud data. More generally, these extensions can be used with any kind of 3D point cloud data which comply with the stochastic CFAR assumptions. The extended CFAR detectors are capable of processing a 2D projected version of the 3D data or they can work directly on the 3D point cloud. The main modifications to the original methods include making the CFAR window size parameter an adaptive one and adding the capability to work with sparse data, in contrast to dense data which is what the original methods expect. The extended CFAR detectors show a more robust performance than other methods when the point cloud data contains high noise and clutter rates. The output of these detector applied to ladar 3D data could then be used for algorithms requiring high accuracy in the detection, for instance, the Iterative Closest Point (ICP) registration method.