Dilated Convolution and Adaptive Upsampling-Based FCN Framework for High-Precision Semantic Segmentation in Self-Driving Vehicles

Autonomous driving requires highly precise, real-time vision systems to ensure safe navigation in dynamic and complex environments. The detection and categorization of neighboring scene objects, such as pedestrians, roadways, and vehicles, is challenging. Semantic segmentation has been performed in previous works. However, they have several weaknesses. They are limited in correctly identifying critical objects like lanes, road boundaries, and pedestrians. Boundary detection and segmentation accuracy are further limited by the loss of spatial information caused by feature fattening. Moreover, when applied to poor lighting or crowded regions, they have deficiencies. Some of these models have increased processing time and memory requirements. To address these limitations, we propose an advanced FCN-based semantic segmentation. This work’s main contributions are that we use dilated convolutions to expand the receptive field without adding more parameters. It made it possible for the model to effectively collect contextual information as well as local characteristics. Next, we refine segmentation results using the adaptive upsampling technique. Even in complicated environments, this helps in accurately identifying boundaries between closely spaced items. Using these integrations, the proposed model yields better results in semantic segmentation without high computation time. Various experiments are conducted to validate the performance of the model using the Lyft Udacity Challenge dataset. The pixel accuracy of 93.04% was obtained with improved performance over several state-of-the-art models. It demonstrates that the proposed FCN-based framework significantly enhances scene understanding for autonomous driving, contributing toward more reliable navigation in real-world environments.