A Robot Localization Deviation Detection Method Based on ResNet

Simultaneous location and mapping (SLAM) based on particle filtering is a technique that uses probabilistic methods to solve the problem of robot localization and mapping in unknown environments. It represents the probability density function of state by using a set of particles, which denotes the probability distribution of the position of the robot and map state. The divergence of particles typically indicates that the estimated pose of the robot is starting to deviate, leading to a decrease in mapping and localization accuracy. In order to detect the degree of particle divergence, we have designed a neural network based on Residual Network (ResNet) architecture specifically for detecting particle distribution. Through supervised learning, this model has been endowed with the ability to analyze particle distribution and determine the particle divergence. To achieve this, we have proposed a dedicated dataset in which we have designed a method to create scale-invariant images based on particle coordinates. The model achieves an accuracy rate of 95.1 % on the test set, and we have also demonstrated its optimality through ablation experiments.