Objective When subway vehicle number recognition detection points are affected by the onsite installation environment, and the camera's mounting position is too close to the vehicle body, it may lead to a limited camera field of view failing to capture the entire subway vehicle number. To address this situation, a recognition method for subway vehicle number images with incomplete character length is proposed.

Method The method mainly consists of two independent thread modules of image acquisition and vehicle number recognition. The former continuously captures subway body images and saves them sequentially to an image queue. While the latter continuously retrieves images from the queue for recognition and then removes them from the queue. Image recognition mainly uses the region object detection model to locate the area of the first two vehicle number digits, the remaining digits area, and the gangway area, and save them to queue 1, queue 2, and queue 3 respectively. The system then determines whether the vehicle separation nod has been reached based on the data volume in these three queues. If the split point has not been reached, the system continues to retrieve images from the image queue. Once the split nod is reached, a character object detection model is used to individually recognize the characters results in queue 1 and queue 2. Finally, the complete subway vehicle number is stitched together using the validation stitching logic, and returned to the backend as the current vehicle number recognition result, then the data in queue 1, queue 2, and queue 3 are cleared.

Result & Conclusion  The actual onsite running test results shows that with this method the recognition rate reaches 100% within one week of trial operation. The recognition speed meets the real-time requirements, allowing for complete vehicle number recognition from images with an incomplete character count, even while the subway is in motion without stopping.