Abstract: Aiming at the problems that the traditional  hierarchical matching algorithm of visual image for mobile robot could only complete rough matching, resulting in low final matching accuracy and long matching time, we proposed a  hierarchical matching algorithm of visual image for mobile robot based on deep reinforcement learning.  Firstly, the strategy network and value network in the deep reinforcement learning network structure were used to guide the floating image to move to the reference image in the correct direction. Secondly, in the process of rough matching, the reward function was designed to realize the rough matching of color features. Finally, on the basis of rough matching, the improved scale invariant feature transformation algorithm was used to extract  the local features of the image to be matched, and the mobile robot visual image was graded matched according to the similarity. The experimental results show that the algorithm can effectively realize the coarse matching and fine matching of images. The stability of feature detection is high under different viewing angles and scales. The matching accuracy is high, the time is short, and the matched image quality is good, which improves the actual application effect of mobile robot.

Key words: deep reinforcement learning, mobile robot, visual image, coarse matching, fine matching, reward function