The invention discloses a deep learning vehicle counting method based on road surface extraction and segmentation. The method specifically comprises: collecting video images of a road by using a camera; adopting a digital image processing method to acquire a road surface area and adopting a segmentation strategy to segment a road surface into a near end part and a far end part; sending the segmented road surface area into a deep learning network to detect a vehicle target, carrying out continuous tracking according to a detection result to obtain a vehicle two-dimensional trajectory, and counting the flow of different types of vehicles in a certain road direction by using the vehicle two-dimensional trajectory to achieve the purpose of vehicle counting. According to the method, the detection precision of small vehicles far away from the road surface is high, and a data basis is provided for accurate vehicle counting. The method can be applied to various traffic scenes, has high stability and counting precision, can effectively and accurately detect and continuously track vehicles in a road range in an image visual field so as to realize vehicle counting, and has a wide applicationprospect.