An embedded vision based docking approach of child robot is proposed for the child robot recovery in the marsupial robotic system . Based on the embedded vision platform, the child robot captures and processes the image data of identifier set on the entrance of the docking station, and extracts the center of guiding identifier fast according to local color distribution features. Then, the horizontal color boundary line is fitted out with the least square method, involving the corresponding imaging slope. By referring to the model of imaging slope-deviation angle of observation, the child robot judges out the relationship between its heading and the normal vector of the entrance to make them basically overlap to each other and finally finish docking. The proposed approach can achieve the accurate docking without camera calibration and its effectiveness is verified by experiments.