This thesis describes the development of a range sensing system. The goal was to create a range sensor that is robust and flexible so that a number of applications within the forest products manufacturing environment can be addressed. Features of the system include: the capability of producing spatially registered image pairs of range and intensity, the ability to generate both range and intensity very quickly, the applicability to a wide variety of industrial applications, the ability to handle large depth-of-field range sensing problems, the ability to do real-time data processing, and the capability to do extensive system diagnostics under complete software control. A triangulation based plane-of-light optical method is employed to extract range information. The research shows that this method suits range sensing applications where conveyor belts are involved. An in-depth study of the triangulation method is included. In the study it shows that this method also supports large depth-of-field range sensing. A dedicated signal processing hardware, built on the Micro Channel interface, performs pipelined image processing and generates range and intensity images in a spatially registered form. The hardware is designed to support several modes of operation, for the purpose of facilitating optical adjustments and calibrations. The hardware self-diagnostic facility is also included in the design. A memory management scheme is provided that facilitates real-time data processing of the range and intensity images. The experiments show that this scheme provides a real-time environment for software processing. This thesis also contains a theory exploring the limitations of the measurement accuracy of the range detection algorithm employed in the prototype system. The maximum data generation rate of the prototype system is 380 range/intensity lines per second at 128 range/intensity pixels per line. Several proposals toward future work are included that aim at improving the speed as well as the measurement accuracy of the prototype system. / Master of Science

A fundamental problem in spacecraft proximity operations is the determination of the 6 degree of freedom relative navigation solution between the observer reference frame and a reference frame tied to a proximal body. For the most unconstrained case, the proximal body may be uncontrolled, and the observer spacecraft has no a priori information on the body. A spacecraft in this scenario must simultaneously map the generally poorly known body being observed, and safely navigate relative to it. Simultaneous localization and mapping(SLAM)is a difficult problem which has been the focus of research in recent years. The most promising approaches extract local features in 2D or 3D measurements and track them in subsequent observations by means of matching a descriptor. These methods exist for both active sensors such as Light Detection and Ranging(LIDAR) or laser RADAR(LADAR), and passive sensors such as CCD and CMOS camera systems. This dissertation presents a method for fusing time of flight(ToF) range data inherent to scanning LIDAR systems with the passive light field measurements of optical systems, extracting features which exploit information from each sensor, and solving the unique SLAM problem inherent to spacecraft proximity operations. Scale Space analysis is extended to unstructured 3D point clouds by means of an approximation to the Laplace Beltrami operator which computes the scale space on a manifold embedded in 3D object space using Gaussian convolutions based on a geodesic distance weighting. The construction of the scale space is shown to be equivalent to both the application of the diffusion equation to the surface data, as well as the surface evolution process which results from mean curvature flow. Geometric features are localized in regions of high spatial curvature or large diffusion displacements at multiple scales. The extracted interest points are associated with a local multi-field descriptor constructed from measured data in the object space. Defining features in object space instead of image space is shown to bean important step making the simultaneous consideration of co-registered texture and the associated geometry possible. These descriptors known as Multi-Field Diffusion Flow Signatures encode the shape, and multi-texture information of local neighborhoods in textured range data. Multi-Field Diffusion Flow Signatures display utility in difficult space scenarios including high contrast and saturating lighting conditions, bland and repeating textures, as well as non-Lambertian surfaces. The effectiveness and utility of Multi-Field Multi-Scale(MFMS) Features described by Multi-Field Diffusion Flow Signatures is evaluated using real data from proximity operation experiments performed at the Land Air and Space Robotics(LASR) Laboratory at Texas A&M University.

This thesis is about the intelligent compression of large 3D point cloud datasets. The non-parametric method that we describe simultaneously generates a continuous representation of the workspace surfaces from discrete laser samples and decimates the dataset, retaining only locally salient samples. Our framework attains decimation factors in excess of two orders of magnitude without significant degradation in fidelity. The work presented here has a specific focus on gathering and processing laser measurements taken from a moving platform in outdoor workspaces. We introduce a somewhat unusual parameterisation of the problem and look to Gaussian Processes as the fundamental machinery in our processing pipeline. Our system compresses laser data in a fashion that is naturally sympathetic to the underlying structure and complexity of the workspace. In geometrically complex areas, compression is lower than that in geometrically bland areas. We focus on this property in detail and it leads us well beyond a simple application of non-parametric techniques. Indeed, towards the end of the thesis we develop a non-stationary GP framework whereby our regression model adapts to the local workspace complexity. Throughout we construct our algorithms so that they may be efficiently implemented. In addition, we present a detailed analysis of the proposed system and investigate model parameters, metric errors and data compression rates. Finally, we note that this work is predicated on a substantial amount of robotics engineering which has allowed us to produce a high quality, peer reviewed, dataset - the first of its kind.

This work is a first step towards a long-term goal of understanding computations occurring in the brain and using those principles to make more efficient machines. The traditional computing paradigm calls for using digital supercomputers to simulate large scale brain-like neural networks resulting in large power consumption which limits scalability or model detail. For example, IBM's digital simulation of a cat brain with simplistic neurons and synapses consumes power equivalent to that of a thousand houses! Instead of digital methods, this work uses analog processing concepts to develop scalable, low-power silicon models of neurons which have been shown to be around ten thousand times more power efficient. This has been achieved by modeling the dynamical behavior of Hodgkin-Huxley (H-H) or Morris-Lecar type equations instead of modeling the exact equations themselves. In particular, the two silicon neuron designs described exhibit a Hopf and a saddle-node bifurcation. Conditions for the bifurcations allow the identification of correct biasing regimes for the neurons. Also, since the hardware neurons compute in real time, they can be used for dynamic clamp protocols in addition to computational experiments. To empower this analog implementation with the flexibility of a digital simulation, a family of field programmable analog array (FPAA) architectures have been developed in 0.35 um CMOS that provide reconfigurability in the network of neurons as well as tunability of individual neuron parameters. This programmability is obtained using floating-gate (FG) transistors. The neurons are organized in blocks called computational analog blocks (CAB) which are embedded in a programmable switch matrix. An unique feature of the architecture is that the switches, being FG elements, can be used also for computation leading to more than 50,000 analog parameters in 9 sq. mm. Several neural systems including central pattern generators and coincidence detectors are demonstrated. Also, a separate chip that is capable of implementing signal processing algorithms has been designed by modifying the CAB elements to include transconductors, multipliers etc. Several systems including an AM demodulator and a speech processor are presented. An important contribution of this work is developing an architecture for programming the FG elements over a wide dynamic range of currents. An adaptive logarithmic transimpedance amplifier is used for this purpose. This design provides a general solution for wide dynamic range current measurement with a low power dissipation and has been used in imaging chips too. A new generation of integrated circuits have also been designed that are 25 sq. mm in area and contain several new features including adaptive synapses and support for smart sensors. These designs and the previous ones should allow prototyping and rapid development of several neurally inspired systems and pave the path for the design of larger and more complex brain like adaptive neural networks.

In this thesis, we present three main contributions to laser-based detection and tracking of dynamic objects, from both a model-based point of view and a model-free point of view, with an emphasis on applications to autonomous driving. A segmentation-based detector is first proposed to provide an end-to-end detection of the classes car, pedestrian and bicyclist in 3D laser data amongst significant background clutter. We postulate that, for the particular classes considered, solving a binary classification task outperforms approaches that tackle the multi-class problem directly. This is confirmed using custom and third-party datasets gathered of urban street scenes. The sliding window approach to object detection, while ubiquitous in the Computer Vision community, is largely neglected in laser-based object detectors, possibly due to its perceived computational inefficiency. We give a second thought to this opinion in this thesis, and demonstrate that, by fully exploiting the sparsity of the problem, exhaustive window searching in 3D can be made efficient. We prove the mathematical equivalence between sparse convolution and voting, and devise an efficient algorithm to compute exactly the detection scores at all window locations, processing a complete Velodyne scan containing 100K points in less than half a second. Its superior performance is demonstrated on the KITTI dataset, and compares commensurably with state of the art vision approaches. A new model-free approach to detection and tracking of moving objects with a 2D lidar is then proposed aiming at detecting dynamic objects of arbitrary shapes and classes. Objects are modelled by a set of rigidly attached sample points along their boundaries whose positions are initialised with and updated by raw laser measurements, allowing a flexible, nonparametric representation. Dealing with raw laser points poses a significant challenge to data association. We propose a hierarchical approach, and present a new variant of the well-known Joint Compatibility Branch and Bound algorithm to handle large numbers of measurements. The system is systematically calibrated on real world data containing 7.5K labelled object examples and validated on 6K test cases. Its performance is demonstrated over an existing industry standard targeted at the same problem domain as well as a classical approach to model-free tracking.