The work aims to solve the problems of low complexity and high repetition rate of laser sensor information for material delivery AGVs in guiding roller production workshops, which often leads to the loss of particles near the correct pose during iterative resampling and results in positioning failures. A visual-based adaptive Monte Carlo localization algorithm was proposed. A camera observation model and an autonomous guided vehicle (AGV) body motion model were established. The observation model was processed for distortion correction to complete camera calibration. A visual-based adaptive Monte Carlo algorithm was designed to extract feature information and classify it using a bag-of-words model. A 2D map was constructed using laser radar, and feature point matching was utilized to estimate the pose and achieve self-localization of the AGV. The simulation experiments showed that, compared with the traditional Adaptive Monte Carlo Localization (AMCL) algorithm, the proposed localization algorithm enabled the robot to quickly converge to a highly accurate pose and exhibited superior localization performance. The visual-based AMCL algorithm is designed to achieve high-precision positioning for AGVs in guiding roller production workshops, optimizing operational processes and enhancing the level of intelligent operation in production line systems. It can provide reference for locating applications in other scenarios.