Segmentación de Superficies Viales Mediante Visión IA

Abstract

Automated Road surface detection from video sequences poses a challenge in computer vision due to environmental interference. In this study, an image segmentation system was developed by combining digital processing techniques and deep learning neural networks to identify the road region for automotive applications. The methodology began with the creation of a dataset comprising 646 images extracted from video, which were normalized to 128 x 128 pixels. Subsequently, a BGR color-thresholding segmentation algorithm was applied to isolate the roadway and generate target masks. Using this data, an encoder-decoder convolutional model was trained over 60 epochs. As a result, the neural network achieved an accuracy exceeding 95% and a reduced validation loss, effectively isolating the road surface while filtering out external elements such as vegetation and vehicles. This workflow establishes a solid foundation for visual perception systems in driver assistance and autonomous navigation.

This work establishes a solid foundation for visual perception and autonomous navigation systems. The results demonstrate that preprocessing via BGR thresholding significantly optimizes the image segmentation stage, enabling the encoding-decoding process to efficiently learn road features despite the compact frame size (128x128 pixels). Furthermore, the stability of the loss function across 60 training epochs confirms the absence of overfitting, validating the model's accuracy for computer vision applications in the automotive sector.

The fact that the loss function converged without oscillations over the 60 epochs confirms the stability and proper initialization of the neural network; this system not only demonstrates a minimal margin of error but also presents a highly viable solution for resource-constrained environments.

This paves the way for future implementations in artificial intelligence systems for real-time video sequence processing that uphold international safety standards for autonomous vehicles.