As wheel-driven intelligent robots, Automated Guided Vehicles (AGVs) are increasingly utilized in warehousing, military, and household fields owing to their intelligence, flexibility, and convenience. In this paper, obstacle detection data through obtaining columnar pixels according to the structure of the AGV system from environmental images are presented. The columnar pixel algorithm, which combines conditional probability models and dynamic programming, is used for obstacle zone recognition and navigable area segmentation with the environmental image data. This allows rapid and accurate obstacle recognition and detection under the condition of parallax. Then, using model-based method (the Kalman filter), the trajectory of obstacles can be predicted, which builds a mathematical motion model to forecast the collision point. It will offer references for adjusting the path of AGVs. Taking into account the principle of AGV obstacle avoidance, the path adjustment strategy applies the concept of global guidance combined with local planning. With the help of the advantage of the Q-learning algorithm in adapting the environment, local path planning strategies are proposed. Through the support of the visual navigation system, the path adjustment strategy in this study can obtain the shortest path planning time with average risk being 0.15.
