Vision-based track-before-detect algorithm for multiple target

Abstract

This thesis deals with the vision-based track-before-detect algorithm for multiple target. The track-before-detect method is for detecting and tracking a target having a very low SNR in an image. The track-before-detect method estimates the probability of a target using multiple observations to detect and track the target. In this thesis, a particle filter that shows good performance to the tracking of long-range targets that appear blurry with low SNR, which is difficult to detect, and to the nonlinearly moving target is applied for track-before-detect. A multi-target particle filter using sub-window for track-before-detect also is introduced in order to detect and track multiple targets as well as single target detection with higher performance. The multi-target particle filter using sub-window for track-before-detect algorithm tracks each detected target that is set, respectively, with the sub-window and finds new targets in the area except the sub-window of the image. Since each target is tracked independently using a particle filter within each sub-window, it is advantageous than the conventional multi target particle filter to track multiple targets without missing even if the SNR of any one of the multiple targets become weak. It also shows robust performance against rapidly changing targets since the size of the sub-window is changed according to the target size change The proposed algorithm is demonstrated to be superior to the conventional multi target particle filter by applying it to the actual captured airplane target video. Monte Carlo simulations were performed to compare the performance of the multi target particle filter using sub-window with the conventional multi target particle filter, in terms of the multi target detection performance, OSPA, and computation time.