In order to improve the visual obstacle avoidance performance of inspection robot

a visual obstacle avoidance method for robot based on scene understanding and probabilistic sampling is proposed.The method constructs a scene understanding network using lightweight depthwise separable dilated convolution units as the core to identify road and obstacle information in the scene.Then

the scene understanding results are transformed into a binarized semantic map

and a heuristic local probabilistic sampling algorithm is designed to efficiently search for safe obstacle avoidance paths from the semantic map.Finally

a feature temporal analysis structure is introduced to filter redundant feature computations during the robot’s inspection process

ensuring the method’s efficiency in practical applications.Experimental results show that the proposed method can effectively enhance the robot’s scene understanding capability and obstacle avoidance performance.Compared with other methods

it also demonstrates stronger robustness and stability

and can better assist the robot in completing the inspection task.