5. Conclusion

The proposed license plate detection model showed the ability to detect license plates under different vision angles. Performance estimation on the dataset with oblique vehicles outperformed the existing commercial systems OpenALPR and SightHound, reached mAP/mAP50 of 40.8/90.1. Our system also detects bounding quadrilateral instead of bounding rectangles, yielding a more precise indication for license plate compared to conventional systems.

Another main contribution is providing the vehicle information while performing license plate detection, we called this kind of information contextual information, which provides the relation comprehension between the license plate and the vehicle, we got the pose classification accuracy 98.8% and average IoU 71.3%. Applications like traffic scene analysis, we may utilize contextual information for enhancing the interpretation of the license plate. Since we have obtained the area of the owner car, by further analyzing, we can get; for instance car brand, model, and color information. In addition, some parking lots tell users to park their cars in a consistent direction (e.g., back-in parking only), the pose information given by our system might help the management of those parking lots.

To accomplish our purpose, we designed a novel anchor-free method for region proposal, which is not limited to the license plate detection task but can be extended to other object detection tasks. The method might be an inspiration for future object detector design.

The model is designed as a one-stage detector, which is fast enough to perform in real-time, meeting the request for most license plate detection applications. But the drawback exists, as a one-stage detector, finding a highly refined bounding area is more difficult than two-stage detectors do.

A dataset with manually annotated vehicle front-rear regions is proposed in our work, which is served as a key element to train our model successfully. Another dataset with extremely oblique but readable license plates is also proposed, with the help of the dataset, generalization of applicable scenes for license plate detection system can be further improved. We are looking forward to broader usages of the datasets for applications related to car pose estimation and license plate detection.

For future works, adding more functional heads might be a direction. Our model is flexible since we can add several functional heads after extracting features from the backbone network. Functions like car brand and car model analysis or even more precise car pose estimation given in rotation angle might be trained. Introducing the OCR into the end-to-end network to make a complete ALPR system is also an effort worthy to put.