Hardware-Accelerated Vehicle License Plate Detection at High-Definition Image

Localizing license plate in an image enables vehicle detection and identification. Processing at high-definition (HD) image allows a better access of visual information but also enhances the necessity of multi-resolution analysis because license plates may appear in various sizes and shapes. A great computational burden is accompanied by processing the great number of candidate windows during multi-resolution analysis. This paper proposes a hardwired license plate detector capable of conducting feature analysis on millions of candidate windows with the size of 100-by-100 pixel in average for one HD image. It employed a multiple-layer classification analyzing color, gradient and texture features at the rate of thirty frames per second. Candidate window fusion subsequent to classification compensates the over-detection occurred during multi-resolution analysis. A dual-loop processing is adopted to schedule the efforts in feature extraction to achieve a real-time system. Compared to the pure software implementation, a third-order speed up was found for the same algorithm with hardwired acceleration in FPGA emulation.