A unimodal degradation detection method for particle filter-based slam algorithms

Abstract

To address the problem that the error of simultaneous location and mapping(SLAM) will increase dramatically in degraded environments that invalidate sensor data. This article introduces a detection method for identifying degraded environments in particle filter-based SLAM algorithms. We develop a Multilayer Perceptron(MLP) that takes particles from Rao-Blackwellized Particle Filter(RBPF) as inputs, enabling it to learn scene degradation features from particle distributions. Typically, the distribution of these particles reflects the localization accuracy, and we let the neural network analyze it, replacing the complex process of manually setting up the features required by the traditional machine learning paradigm. The model achieved 90% accuracy on the test set, demonstrating its ability to recognize degraded scenes. We deployed the model on a robot, integrating it with SLAM via Robot Operating System(ROS) for real-time environmental sensing. Although accuracy dipped in new scenarios, the potential of the proposed method for future SLAM applications remains promising.