Traditional vehicle recognition methods have the disadvantages such as low efficiency and time-consuming due to the complex background and overlapping situation. In this paper, we propose a multiscale adptive spatial feature fusion (ASFF) method for vehicle recognition. First, it calculates the difference hash values of images. Then the hash value is used to judge the similarity between the current frame and the previous frame. When the similarity is less than the threshold value, it is input to ResNet18 model for detection. Using ResNet18 as the base network can reduce network parameters. Then, aiming at the problem that the detection effect of ASFF for vehicle recognition is not ideal, the offset loss and width-height loss are replaced by the intersection ratio loss. Meanwhile, multi-scale adaptive spatial feature fusion method is adopted to fuse the multi-level features of the network. The experimental results show that the average accuracy with proposed methed is increased by 2.1%. For BDD100K and Pascal VOC datasets, the average accuracy of predicted borders is increased by 5.5%, when the IoU is greater than 0.5. With the GTX1060Ti, the recognition speed can reach 149 frames per second. The multiscale ASFF in this paper can significantly improve the vehicle recognition accuracy.