During the image local feature matching process，error matches will be eliminated effectively by conside-ring the motion statistics of features． However，the current grid-based method of motion statistics works poorly with zoom and rotation． To solve this problem，a screening method for feature matching based on dynamic window mo-tion statistics was proposed． Firstly，the algorithm builds a fast approximate nearest neighbor index structure based on the location of image feature points． Then it sets up the dynamic window and computes motion statistics． Fina-lly，it eliminates error matches with the score of motion statistics． The experimental results show that，compared with other methods，the proposed method has a significant advantage over the algorithm based on grid in predicating precision and recall rate when the scale and angle change greatly． And in more general scenarios，the overall matc-hing effect of this algorithm is better than other real-time matching methods． Meanwhile，this algorithm has good time performance and can be applied to real-time tasks．