The invention relates to an autonomous underwater robot collision avoidance method in an unknown complex seabed terrain environment. The method comprises the following steps that an underwater robot processes collected sonar image data and acquires an obstacle object; and according to an area position of the obstacle object in a sonar image, determining an AUV collision avoidance behavior in realtime. In the invention, an environmental-perception false alarm rate is low. A foresight sonar image is acquired, an image processing mode is used to perceive an obstacle and accuracy of obstacle identification is increased. An AUV can safely execute a task without collision in the complex seabed terrain environment. The AUV is ensured not to generate near-bottom collision and a bottom loss problem is not generated either. Detection efficiency is increased. A track requirement of AUV detection work is guaranteed, and conditions of repeated covering or white spaces are less.