RANSAC-Based Track Initiation for Kalman Filtering and Multi-Hypothesis Tracking Systems in Cluttered Range-Only Environments

Abstract

Track initiation in cluttered range-time surveillance environments is a challenging problem. Classical M-of-N initiation methods confirm tracks based on detection count alone within a static gate, with no geometric validation, making them susceptible to false track initialization as clutter density increases. This thesis proposes a RANSAC (Random Sample Consensus) based track incubation framework that utilizes the linear geometry of constant velocity targets in the range-time plane to differentiate genuine tracks from clutter before promotion. Candidate tracks are confirmed only when a sufficient number of detections support a linear hypothesis under random sampling and promoted tracks are initialized using a Weighted Least Squares fit that provides range, range-rate, and covariance. The incubation framework is evaluated in combination with two engines, a Global Nearest Neighbor Kalman Filter and a Track-Oriented Multiple Hypothesis Tracker, and benchmarked against M-of-N initialized counterparts across three simulated collision and debris scenarios under nominal, heavy clutter, and reduced detection probability conditions. While both M-of-N engines collapsed to negative MOTA under heavy clutter, the RANSAC-TOMHT engine maintained a positive MOTA (Multiple Object Tracking Accuracy). The RANSAC-GNNKF engine produced a MOTA that was orders of magnitude higher than M-of-N. This result demonstrates that geometric validation at initiation is beneficial for reliable tracking in high clutter environments. RANSAC engines outperformed M-of-N counterparts on MOTA, MOTP, and RMSE, at the cost of increased runtime. A Monte Carlo validation study confirmed that the WLS covariance is theoretically sound but subject to overconfidence due to inlier set contamination in clutter, an inherent limitation of RANSAC-based initiation.