Automated detection of abnormal driving events is crucial for improving road safety and emergency response. Current detection systems combine inertial sensors, cameras, and neural networks to identify risky driving behaviors. However, most approaches either require costly full-scale infrastructure or depend on visual systems and complex network architectures. This study presents a behavior-based event detection framework that combines short-time Fourier transform with support vector machine (SVM) classification. Using a scaled-vehicle that accurately replicates real driving conditions, we analyzed five characteristic events: deformable-object impact, progressive acceleration, constant velocity, sudden braking, and speed bump overpass. Our methodology extracts distinctive time-frequency patterns from triaxial acceleration data, which serve as input features for a soft-margin SVM classifier with radial basis function kernel. The system achieved a classification performance of 96.63% accuracy, which indicates the feasibility of this approach for identifying driver events. These results suggest potential applications in vehicle safety systems and driver assistance technologies, although further validation at full scale is required.