SBMEV: A Stacking-Based Meta-Ensemble Vehicle Classification Framework for Real-World Traffic Surveillance

Developing vehicle classification remains a fundamental challenge for intelligent traffic management in the Indian urban environment, where traffic exhibits high heterogeneity, density and unpredictability. In the Indian subcontinent, vehicle movement is erratic, congestion is high, and vehicle types vary significantly. Conventional global benchmarks often fail to capture these complexities, highlighting the need for a region-specific dataset. To address this gap, the present study introduced the EAHVSD dataset, a novel real-world image collection comprising 10,864 vehicle images from four distinct classes, acquired from roadside surveillance cameras at multiple viewpoints and under varying conditions. This dataset is designed to support the development of an automatic traffic counter and classifier (ATCC) system. A comprehensive evaluation of eleven state-of-the-art deep learning models, namely VGG16, VGG19, MobileNetV2, Xception, AlexNet, ResNet50, ResNet152, DenseNet121, DenseNet201, InceptionV3, and NASNetMobile, was carried out. Among these, the highest accuracy result has been achieved by VGG-16, MobileNetV2, InceptionV3, DenseNet-121, and DenseNet-201. We developed a stacking-based meta-ensemble framework to leverage the complementary strengths of its components and overcome their individual limitations. In this approach, a meta-learner classifier integrates the predictions of the best-performing models, thereby improving robustness, scalability, and real-world adaptability. The proposed ensemble model achieved an overall classification accuracy of 96.04%, a Cohen’s Kappa of 0.93, and an AUC of 0.99, consistently outperforming the individual models and existing baselines. A comparative analysis with prior studies further validates the efficacy and reliability of the stacking-based meta-ensemble method. These findings position the proposed frameworks as a robust and scalable solution for efficient vehicle classification under practical surveillance constraints, with potential applications in intelligent transportation systems and traffic management.