In order to achieve the attitude adjustment of the unmanned car in no-mark maze

on the basis of fusion of ultrasonic sensor and camera data

a miniature fully convolutional networks（ FCN） is proposed to divide the ground in real-time

detect the maze lane line and for precise adjustment. Compared with the common FCN

the parameter volume of the network drops by 97 %

the memory usage drops by 86. 44 %

and the detection frame rate reaches 11. 2 fps. So it can be used for embedded systems carrying. After the actual test

the algorithm has good adjustment effect

and can be used for unmanned car attitude adjustment in real-time in maze.