International Journal of Science, Engineering and TechnologyAuthors: Roshan P
Abstract: Vehicular safety on public roads stands as a pressing societal concern worldwide, with pavement craters ranking among the foremost contributors to traffic collisions and automobile damage. An autonomous pothole recognition framework has been engineered and deployed, employing the YOLOv8 deep learning architecture to locate pavement voids in continuous video streams acquired from onboard cameras. Capable of analyzing footage from dashboard-mounted cameras or handheld devices, the framework delivers instantaneous audible hazard notifications to vehicle operators upon each confirmed detection event. The architectural design integrates a browser-accessible front end developed with HTML, CSS, and Bootstrap, while Python augmented by the Flask microframework governs all computational processing and model execution at the server side. The YOLOv8 detection engine was trained on a meticulously curated, annotated corpus spanning heterogeneous road surfaces, illumination regimes, and meteorological conditions, ultimately attaining a mean Average Precision (mAP@0.5) exceeding 89% on withheld evaluation data. Empirical assessments demonstrate that the framework operates with sustained reliability, sustaining throughput at a mean rate of 40 frames per second on commodity computing platforms. Extending beyond individual driver protection, the framework simultaneously archives detection telemetry for use by pavement management agencies in scheduling remediation campaigns. The work illustrates how operationally deployable deep learning technologies can advance the aspirations of smarter road governance and more resilient transport networks. Specifically, the framework attains a precision of 94.2% under unobstructed daytime lighting and an aggregate mAP@0.5 of 89.1% across all evaluation scenarios, accompanied by a mean inference throughput of 40.3 frames per second, corroborating its readiness for real-time driver-assist deployment on widely accessible consumer platforms.
DOI: http://doi.org/10.5281/zenodo.20721532
