Abstract: Deep convolutional neural networks(DCNN)have been widely adopted in visual object detection as they can automatically learn effective image features and have strong generalization capabilities.Since the current mainstream visual target detection DCNN commonly output rectangular bounding boxes,they cannot effectively deal with the accurate location for tilted unconstrained license plates.To solve the problem,a solution was proposed and YOLOv3,as a specific DCNN,was trained by extending its output dimensions and establishing additional loss items about vertex offsets relative to the four bounding box corners,which empowers the trained model to output the rectangular bounding boxes and the license plate vertices meanwhile without sacrificing its high-efficiency performance.Experiments on CCPD,a widely used large scale unconstrained license plate dataset,illustrate that the proposed method can not only accurately predict the four vertices of a license plate,but also reach the precision above 99%in Base,Tilt and Weather subset.Moreover,this method can also be generalized to other application fields which require the target bounding box and key points output simultaneously and has promising application potential.