Theses and Dissertations from UMD
Permanent URI for this communityhttp://hdl.handle.net/1903/2
New submissions to the thesis/dissertation collections are added automatically as they are received from the Graduate School. Currently, the Graduate School deposits all theses and dissertations from a given semester after the official graduation date. This means that there may be up to a 4 month delay in the appearance of a give thesis/dissertation in DRUM
More information is available at Theses and Dissertations at University of Maryland Libraries.
Browse
Search Results
Item The Curved Openspace Algorithm and Neuromorphic Mechanisms for Sonar-Based Obstacle Avoidance(2021) Wen, Chenxi; Horiuchi, Timothy TKH; Electrical Engineering; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)Bats are known for their ability to pursue a goal while avoiding obstacles in a cluttered environment using ultrasonic echolocation. This dissertation explores two neuromorphic mechanisms involved in such a task: a conductance-based neuron circuit for azimuthal echolocation whose dynamic range can be expanded by power-law compression, and a sonar-based obstacle avoidance algorithm with its spike-latency model implementation. Bats and other mammals use interaural level differences (ILD) to estimate the direction of high-frequency sounds. To compute the ILD of a sound, independent of overall loudness, excitatory and inhibitory synaptic conductances (encoding the left and right amplitudes) are hypothesized to compete in the neurons of the lateral superior olive. Interestingly, this neural model can also accept power-law compressed amplitudes that can allow a much larger range of input signal levels, a common limitation in neural coding. This dissertation demonstrates the use of square-root and cube-root compression with a neuromorphic VLSI neuron to expand the range of distances over which ILD can be used to estimate echo direction in a sonar system based on echolocating bats. However, many questions remain regarding how to achieve the rapid control of a sonar-guided vehicle to pursue a goal while avoiding obstacles. Taking into account the limited field-of-view of practical sonar systems and vehicle kinematics, we propose an obstacle avoidance algorithm that maps the 2-D sensory space into a 1-D motor space and evaluates motor actions while combining obstacles and goal information. A winner-take-all (WTA) mechanism is used to select the final steering action. To avoid unnecessary scanning of the environment, an attentional system is proposed to control the directions of sonar pings for efficient, task-driven, sensory data collection. A mobile robot driven by the proposed algorithm was capable of navigating through a cluttered environment using a realistic sonar system. The algorithm was tested on a mobile robot, and it is implemented on a neural model using spike-timing representations, a spike-latency memory, and a “race-to-first-spike” WTA circuit. This dissertation also proposed a CMOS floating-gate circuit for artificial neural network synapse memories that can achieve a fixed rate of weight increase (adding vectors) or proportional decay (normalization) on the synaptic weight.Item BIO-INSPIRED SONAR IN COMPLEX ENVIRONMENTS: ATTENTIVE TRACKING AND VIEW RECOGNITION(2021) Isbell, Jacob D; Horiuchi, Timothy K; Electrical Engineering; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)Bats are known for their unique ability to sense the world through echolocation. This allows them to perceive the world in a way that few animals do, but not without some difficulties. This dissertation explores two such tasks using a bio-inspired sonar system: tracking a target object in cluttered environments, and echo view recognition. The use of echolocation for navigating in dense, cluttered environments can be a challenge due to the need for rapid sampling of nearby objects in the face of delayed echoes from distant objects. If long-delay echoes from a distant object are received after the next pulse is sent out, these “aliased” echoes appear as close-range phantom objects. This dissertation presents three reactive strategies for a high pulse-rate sonar system to combat aliased echoes: (1) changing the interpulse interval to move the aliased echoes away in time from the tracked target, (2) changing positions to create a geometry without aliasing, and (3) a phase-based, transmission beam-shaping strategy to illuminate the target and not the aliasing object. While this task relates to immediate sensing needs and lower level motor loops, view recognition is involved in higher level navigation and planning. Neurons in the mammalian brain (specifically in the hippocampus formation) named “place cells” are thought to reflect this recognition of place and are involved in implementing a spatial map that can be used for path planning and memory recall. We propose hypothetical “echo view cells” that could contribute (along with odometry) to the creation of place cell representations actually observed in bats. We strive to recognize views over extended regions that are many body lengths in size, reducing the number of places to be remembered for a map. We have successfully demonstrated some of this spatial invariance by training feed-forward neural networks (traditional neural networks and spiking neural networks) to recognize 66 distinct places in a laboratory environment over a limited range of translations and rotations. We further show how the echo view cells respond in between known places and how the population of cell outputs can be combined over time for continuity.