Hybrid SVM–CNN Framework for Obstacle Classification in Smart Driving Assistance Systems for Autonomous Vehicles

Accurate obstacle classification is a necessity in autonomous vehicles and intelligent driver aid systems. Current deep learning models, although proficient in feature extraction, frequently demonstrates difficulties in establishing accurate decision boundaries, particularly in settings characterised by varied object kinds and detailed backgrounds. The proposed article offers a Hybrid SVM–CNN Framework that combines Convolutional Neural Networks (CNN) for deep feature extraction with Support Vector Machines (SVM) for high-margin classification to tackle this difficulty. The system is analysed using documented urban road problem datasets obtained from Kaggle, which include automobiles, pedestrians, and road obstacles. The CNN derives layered visual characteristics, which are subsequently transmitted to the SVM for conclusive categorisation. The comparative analysis with baseline models VGG16, ResNet50, and YOLOv5 demonstrates the suggested method's higher performance. The hybrid model attains an accuracy of 98.47%, precision of 97.92%, recall of 98.61%, F1-score of 98.26%, and AUC-ROC of 0.992. These findings confirm the system's reliability and feasibility for real-time application in self-navigating contexts. The suggested approach limits misclassification and offers a scalable basis for future advancements in smart transportation systems.