A Hybrid PSO–GA Based Clustered Sorted Weighting Lookahead Particle Filter For Track-Before-Detect Application

Abstract

In the Track Before Detection (TBD) approach, a target is monitored in an environment with a low signal-to-noise ratio before it is officially identified as a target. TBD is used in various areas such as air traffic control, maritime monitoring, tracking of underwater autonomous vehicles (UAVs), and detecting small objects in remote sensing. A key technique used in TBD is the particle filter (PF). In situations that are nonlinear and non-Gaussian, the PF is a leading method for detecting targets. However, the conventional Particle Filter algorithm suffers from a loss of particle diversity due to degradation and resampling processes, which prevents the particle set from accurately representing the true state probability density function. The Particle Swarm Optimization (PSO) algorithm can help reduce particle degradation in PF, but its effectiveness is heavily influenced by the variance of measurement noise and it is prone to getting stuck in local optima, limiting its filtering precision. To address these issues, a hybrid method combining the Genetic Algorithm (GA) with PSO is proposed. By combining PSO’s fast convergence with GA’s strong global search abilities, this approach improves particle diversity, preserves the effectiveness of high-quality particles, and enhances overall accuracy. GA employs global search operations such as selection, crossover, and mutation. In this paper, we propose a selection method called the clustered sorted weighting lookahead particle filter. This algorithm divides the particle set into several clusters using the k-means clustering algorithm. Within each cluster, particles located in high-importance regions are identified and replicated to replace those in low-importance regions. After this replacement, the clusters are merged using cluster weights. This approach integrates clustering with selective likelihood-based resampling to preserve multiple posterior modes, maintain particle diversity, and enhance root mean square error. The tracking performance of the proposed method is assessed using two nonlinear models.