OptimRSEG: An Optimized Semantic Road Segmentation Model

Abstract

The traditional methods used in road detection for autonomous vehicle applications depend largely on lane marking detection. The techniques can be compromised by shadows and vehicles, occluding the important features crucial to lane detection. This problem is even more prevalent in the case of unstructured roads without markings or borders. This study demonstrated a Particle Swarm optimization (PSO) optimized lightweight semantic segmentation model that made use of AlexNet architecture as its backbone for detecting urban roads, both with and without markings, and under different occlusion conditions. The PSO method is used to search for the best hyperparameters setting to optimize the model learning process using a small dataset for the two-class problem (lane vs. background). Our results showed that the proposed OptimRSEG model produced considerably good performance metrics results of 0.85, 0.91, and 0.923 in the evaluated Intersection of Union (IU), Dice Similarity Coefficient (DSC), and prediction accuracy, respectively. The use of augmentation to enrich the dataset improves these results slightly by around 1-7 %, confirming the effectiveness of the optimization strategy. This system performs acceptably well, even on road images without lane markings or unique markings or partially occluded, with a fast-computing time of 20 ms each frame.