New techniques for image de-noising, thresholding, object detection and their application to vision based collision avoidance system

This dissertation focuses on developing new techniques for image de-noising, thresholding, object detection and their application to collision avoidance system (CAS). Five key research topics presented in this dissertation are: (1) Automatic image noise identification for image de-noising, (2) Adaptive thresholdings for edge and corner detections, (3) Motion based vehicle detection in a vision based CAS system, (4) Vehicle detection based on multiple image features, (5) Target distance estimation and de-noising using interacting multiple model (IMM) based algorithm to find improved estimate of time-to-collision (TTC).;In the area of image de-noising, this dissertation presents a novel noise identification algorithm based on statistical pattern recognition. Noise classification is carried out using three different discriminant functions, which include a newly proposed similarity function, feature weighting (FW) function, and rotated coordinate system (RCS) function. Combining this noise identification algorithm with an appropriate filter results in good image de-noising.;This dissertation also proposes an adaptive thresholding technique for edge and corner detection using mixture model (MM) and maximum likelihood estimation (MLE) techniques. Specifically, Rayleigh mixture model is used to describe the probability distribution of edge strengths, and Gaussian mixture model is used for corner strengths. Experiments on images of varying brightness have produced satisfactory results with edge and corner detection. Detecting front vehicle and estimating the distance between front and host vehicles are two important functions in a rear-end CAS system. This dissertation presents a vehicle detection and distance estimation algorithm based on multiple image features and perspective projection. The experimental results indicate that the proposed algorithm is capable of detecting front vehicle and estimating its distance for rear-end CAS application under normal road conditions.;Time-to-collision (TTC) is calculated by computing the ratio of distance to relative velocity between front and host vehicles. To improve TTC estimates, it is necessary to remove noise from the distance estimates. Assuming that the distance estimates obey either constant velocity (CV) or constant acceleration (CA) model, an interacting multiple model (IMM) algorithm is used to dynamically merge the distance and velocity estimates from CV- and CA-based Kalman filters. Experimental results indicate that the proposed IMM approach works very well to de-noise the distance estimates and thereby improve the estimated TTC.