Kernel Attention SVM for Intelligent Traffic Management in Smart Cities

Abstract

Smart cities increasingly rely on intelligent traffic management systems (ITMS) to address congestion, pollution, and road safety. Machine learning techniques offer promising solutions by enabling real-time traffic pattern recognition and dynamic control strategies. Conventional support vector machines (SVM) often underperform in dynamically changing traffic environments due to limited contextual awareness and generalization capabilities. Deep learning methods, though powerful, may require high computational resources unsuitable for edge devices. This paper proposes a Kernel Attention SVM (KA-SVM) approach, integrating an attention mechanism with traditional SVM to enhance feature representation and improve adaptability to traffic dynamics. The model applies attention over kernel-transformed features to prioritize spatial-temporal patterns in traffic data. Real-time video feeds and sensor signals are processed to detect congestion levels and vehicle flows. The method was trained and tested on the CityFlow dataset using a hybrid simulation environment combining SUMO (Simulation of Urban Mobility) and OpenCV. The KA-SVM achieved a segmentation accuracy of 91.2% and vehicle detection precision of 88.5%, outperforming traditional SVM (79.3%), CNN (85.6%), YOLOv5 (87.1%), and ResNet-SVM hybrid (86.0%). The model also demonstrated robust performance under varying lighting and traffic densities.