Statistical modeling and data fusion of automotive sensors for object detection applications in a driving environment
Abstract
In this work, we consider the application of classical statistical inference to the fusion of data from different sensing technologies for object detection applications in order to increase the overall performance for a given active safety automotive system. Research evolved mainly around a centralized sensor fusion architecture assuming that three non-identical sensors, modeled by corresponding probability density functions (pdfs), provide discrete information of target being present or absent with associated probabilities of detection and false alarm for the sensor fusion engine. The underlying sensing technologies are the following standard automotive sensors: 24.5 GHz radar, high dynamic range infrared camera and a laser-radar. A complete mathematical framework was developed to select the optimal decision rule based on a generalized multinomial distribution resulting from a sum of weighted Bernoulli random variables from the Neyman-Pearson lemma and the likelihood ratio test. Moreover, to better understand the model and to obtain upper bounds on the performance of the fusion rules, we assumed exponential pdfs for each sensor and a parallel mathematical expression was obtained based on a generalized gamma distribution resulting from a sum of weighted exponential random variables for the situation when the continuous random vector of information is available. Mathematical expressions and results were obtained for modeling the following case scenarios: (i) non-identical sensors, (ii) identical sensors, (iii) combination of nonidentical and identical sensors, (iv) faulty sensor operation, (v) dominant sensor operation, (vi) negative sensor operation, and (vii) distributed sensor fusion. The second and final part of this research focused on: (a) simulation of statistical models for each sensing technology, (b) comparisons with distributed fusion, (c) overview of dynamic sensor fusion and adaptive decision rules.
Degree
Ph.D.
Advisors
Bell, Purdue University.
Subject Area
Applied Mathematics|Automotive engineering|Remote sensing
Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server.