Abstract
Intelligent vehicle driving performance is safe and stable, which can significantly improve the efficiency of road traffic and reduce energy consumption, and intelligent vehicle is also the development direction of modern transport. Its core technology is intelligent environment perception module, by using a variety of sensors on the car in which the surrounding environment for data collection, processing module to provide effective control for the basis. In this paper, a new SINS / CNS / GPS integrated navigation observation equation is proposed, and a new federated data fusion structure is designed for the integrated navigation system. The particle filter is used to fuse the multi-source data of the federated filter subsystem, thus eliminating the limitations of the classical Kalman filter. The traditional Kalman filter structure and the federal particle filter mechanism are designed. The comparison shows that the proposed algorithm is effective in the information fusion of the integrated navigation system, and the filtering effect is superior to the traditional filtering method.
Introduction
Navigating an unmanned vehicle in a complex environment requires high-precision positioning and intelligent path planning. Traditionally, researchers have relied on the Kalman Filter (KF) or its variant, the Extended Kalman Filter (EKF), to fuse sensor data. However, these methods have two major Achilles' heels: linearity/Gaussian assumptions and a tendency to get stuck in local optima. This paper addresses these hurdles by using a multi-sensor fusion approach that can handle the complexities of real-world environments through a Federated Particle Filter and an Improved Particle Swarm Optimization (PSO) algorithm.
The Problem with Traditional Filtering
Standard navigation solutions frequently rely on the Kalman Filter for state estimation. As shown in the algorithm loop below, the KF relies on iterative gain calculations and updates that assume systems are linear and noise is Gaussian.
Real-world navigation is often non-linear and "messy," leading to filter divergence in traditional models. Furthermore, standard path-planning algorithms often find paths that look good locally but are inefficient globally.
Proposed Methodology
1. Federated Particle Filter for Navigation
Instead of a single filter, the system uses a federated structure combining a Strapdown Inertial Navigation System (SINS), a Celestial Navigation System (CNS), and GPS.
By replacing the traditional Kalman filters in this architecture with Particle Filters (PF), the system can handle non-linear dynamics without needing cumbersome Jacobian matrices. The federated design provides fault tolerance; if one sensor (like GPS in a tunnel) fails, the system remains stable through global feedback.
2. Improved Particle Swarm Optimization (PSO)
The authors apply an improved PSO to find the best path using a multi-objective fitness function that balances three factors:
- Path Length: Keeping the route as short as possible.
- Safety: A penalty function for proximity to obstacles.
- Smoothness: Accounting for the robot's rotation angles to ensure the path is physically feasible and energy-efficient.
Key Results
In Matlab simulations comparing the "Agile PSO" approach against a traditional Kalman Filter-based path planner, the proposed method showed significant improvements:
| Metric | Kalman Filter (Traditional) | Federated Particle Swarm (Proposed) |
|---|---|---|
| Optimal Path Length | 35.71 | 25.36 |
| Mean Path Length | 42.39 | 33.43 |
| Mean Simulation Time | 0.28s | 0.21s |
Qualitatively, the proposed algorithm demonstrated a superior ability to "jump out" of local optima, finding more direct routes through obstacle clusters where traditional approaches took wider, less efficient detours.
Conclusion
By bypassing the linearization errors of traditional filters and incorporating physical constraints into path planning, the authors achieve more accurate state estimation and safer, smoother paths. Future work will focus on moving from simulation to real-world hardware implementation and expanding the model to 3D environments.
Summary updated