The invention discloses a YOLOv3-based multi-lane traffic flow counting and vehicle tracking method. According to the method, features of an inputted image are extracted by a feature extraction network and an image location and a category probability value are predicted; two adjacent frames of detected vehicle locations are compared and whether the center pointer of the frame of vehicle marking frame is located in a last frame of vehicle marking frame is determined to determine whether the vehicles are the same one, thereby realizing tracking; and with a set detection line, a traffic flow of each lane is obtained based on a line-frame position relationship. Therefore, tracking of the vehicle and vehicle counting at any lane can be realized; the detection speed and the accuracy are also greatly improved; and a traffic accident can be detected.