This thesis examines the performance of some dim target detection algorithms in low-SNR imaging scenarios. In the past research, there have been numerous attempts for detection and tracking barely visible targets for military surveillance applications with infrared sensors. In this work, two of these algorithms are analyzed via extensive simulations. In one of these approaches, dynamic programming is exploited to coherently integrate the visible energy of dim targets over possible relative directions, whereas the other method is a Bayesian formulation for which the target likelihood is updated along time to be able to detect a target moving in any direction. Extensive experiments are conducted for these methods by using synthetic image sequences, as well as some real test data. The simulation results indicate that it is possible to detect dim targets in quite low-SNR conditions. Moreover, the performance might further increase, in case of incorporating any a priori information about the target trajectory.

In this study Track Before Detect (TBD) approach will be analysed for tracking of dim moving targets. First, a radar setup is presented in order to introduce the radar range equation and signal models. Then, preliminary information is given about particle filters. As the main algorithm of this thesis, a multi-model particle filter method is developed in order to solve the non-linear non-Gaussian Bayesian estimation problem. Probability of target existence and RMS estimation accuracy are defined as the performance parameters of the algorithm for very low SNR targets. Simulation results are provided and performance analysis is presented as a conclusion.

No description available.

In this thesis, an algorithm for visual detecting and tracking of very low SNR targets, i.e. dim targets, is developed. Image processing of single frame in time cannot be used for this aim due to the closeness of intensity spectrums of the background and target. Therefore / change detection of super pixels, a group of pixels that has sufficient statistics for likelihood ratio testing, is proposed. Super pixels that are determined as transition points are signed on a binary difference matrix and grouped by 4-Connected Labeling method. Each label is processed to find its vector movement in the next frame by Label Destruction and Centroids Mapping techniques. Candidate centroids are put into Distribution Density Function Maximization and Maximum Histogram Size Filtering methods to find the target related motion vectors. Noise related mappings are eliminated by Range and Maneuver Filtering. Geometrical centroids obtained on each frame are used as the observed target path which is put into Optimum Decoding Based Smoothing Algorithm to smooth and estimate the real target path. Optimum Decoding Based Smoothing Algorithm is based on quantization of possible states, i.e. observed target path centroids, and Viterbi Algorithm. According to the system and observation models, metric values of all possible target paths are computed using observation and transition probabilities. The path which results in maximum metric value at the last frame is decided as the estimated target path.