In order to accelerate the landing of the autonomous valet parking system, a novel autonomous valet parking path planning method is proposed based on the improved fast marching tree algorithm. Firstly, a breadth-first-search-like strategy is used to establish the "path field" of the environmental map, and an obstacle avoidance detection strategy with high computational efficiency is proposed. A selection principle of the far and near reference points and an update principle of the path node based on the "path field" of environmental map are proposed to conform to the vehicle non-holonomic constraint. According to the above proposed strategies and principles, the path node is gradually close to the target node to complete the autonomous valet parking guidance path planning task. Then, on the basis of the Dubins curve, the parking path which meets the arbitrary requirement of the initial vehicle parking azimuth and the non-uniqueness requirement of the parking space azimuth angle is planned to guide the vehicle to enter the parking space safely. Finally, the feasibility of the proposed method is verified by simulation, and the results show that when compared with the traditional fast marching algorithm, the planned autonomous valet parking path based on the proposed method meets the requirement of the vehicle non-holonomic constraint and can guide the vehicle to complete the task of autonomous valet parking.