Voting-based Multi-Plate Recognition for License Plates
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing automatic license plate recognition systems face challenges such as poor file resolution, blurry images, poor lighting, object obstruction, non-standard fonts, and limited real-time tracking capabilities due to reliance on single frame analysis and cloud-based systems, which are hindered by network performance issues.
Innovation Solution
The implementation of a camera system that continually captures video, using a vision processor or specialized circuitry to execute a neural network model for real-time license plate detection and recognition, aggregating frames for improved character detection and optical character recognition, and employing a probing algorithm to optimize CPU usage and precision.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If single frame analysis methods are used for license plate recognition, then the system complexity is reduced, but the tracking capability and recognition accuracy deteriorate
Solution Approach 1:
The system divides the video stream into multiple discrete frames for analysis, processing each frame independently through the neural network model. This segmentation allows the system to maintain simplicity in individual frame processing while achieving robust tracking through temporal aggregation of multiple frame results
Solution Approach 2:
The system continuously processes video frames in real-time rather than relying on periodic or trigger-based analysis. The neural network model executes on every frame (or selected frames) to maintain continuous tracking of license plates, ensuring no movement is missed while keeping the system architecture relatively simple
2Power
If cloud-based systems are used for license plate recognition, then computational power is improved, but network performance issues limit image quality and analysis speed
Solution Approach 1:
The patent introduces a vision processor unit (VPU) as an intermediary between the camera and the cloud system. The VPU performs preliminary neural network inference and frame aggregation locally, processing video frames on the edge device before potentially transmitting results to the cloud. This intermediary approach maintains high image quality by avoiding repeated cloud-uploads of full-resolution frames while still leveraging cloud computational power when needed
Solution Approach 2:
The system replaces the traditional mechanical approach of uploading complete high-resolution images to the cloud with a computational substitution using neural network models that run locally on the vision processor. This substitution allows the system to extract essential features and perform recognition directly on the device, eliminating the need for continuous high-bandwidth image transmission to the cloud
3Measurement precision
If multiple frames are aggregated for license plate recognition, then recognition accuracy is improved, but processing time and computational load increase
Solution Approach 1:
The system aggregates only a limited number of consecutive frames (e.g., 3-5 frames) rather than processing indefinitely many frames. This partial action approach provides sufficient temporal redundancy to overcome single-frame issues like motion blur or occlusion while preventing excessive processing delays. The neural network model processes this manageable subset of frames to achieve improved accuracy without overwhelming computational requirements
Solution Approach 2:
The system performs preliminary processing on individual frames before aggregation, using the neural network model to detect and localize license plates in each frame independently. This preliminary action identifies candidate regions of interest in advance, allowing the aggregation process to focus computational resources only on relevant frames and regions, thereby reducing overall processing time while maintaining accuracy benefits from multi-frame analysis
Data Source
AI summary
A method and related system operations includes, determining, for each respective frame of a frame sequence, a respective bounding box that surrounds a respective sub-image of the respective frame, determining a respective string and respective confidence values associated with the respective string based on the respective sub-image, updating the tracklet to comprise the respective bounding box based on the respective string and at least one string generated by the object recognition model for a previous frame, and updating a voting table by adding the respective confidence values to the voting table. The method also includes generating an aggregated string based on the voting table by, for a set of positions of the aggregated string, determining a character associated with a maximum confidence value indicated by the voting table and associating the aggregated string with the tracklet in a data structure.