Desarrollo de un modulo de visión por computadora para reconocimiento de placas vehiculares en android

Abstract

Automatic vehicle license plate recognition faces challenges when monitoring traffic in cities, especially when users rely on applications to report specific incidents, as the system must be both accurate and efficient. Therefore, this project proposes the development of a module for mobile devices that enables the capture of license plates from previously recorded videos. Furthermore, through iterative improvements to the project, it is expected that identifying the correct license plate will become easier, helping incident reporters save time by providing clear and reliable visual evidence.

A native module in C++ was developed and integrated into an existing Android application using JNI, applying computer vision techniques such as aspect ratio analysis, edge density, Local Binary Patterns (LBP), HSV color analysis, gradients, and SIFT keypoint detection. In addition, urban traffic videos were used, employing OpenCV for frame extraction and Roboflow for dataset annotation, along with performance and security optimizations implemented through CMake.

The results show an individual frame accuracy of 27.05%, with a processing time of 4.3 ms per frame and a cumulative detection probability of 95% across 15 frames. Moreover, user evaluation reported a usability score of 4.2/5 and confidence in the results of 3.8/5. Finally, the system was integrated into the CTD-ESPOL application, reducing operational workload, although further testing in real-world scenarios is still required. The projected development cost was USD 2,400, with annual operational costs of USD 2,160, suggesting long-term feasibility. Keywords: Optimization, Video analysis, Automatic detection, Mobile Interface, Crowdsourcing.