Visual Classification of Vehicle Types Using Transfer Learning

Abstract

Road traffic crashes remain a major cause of death among children and young people, and reliable visual recognition of the vehicles involved can support later accident documentation. Vehicle type classification is well studied on ordinary traffic imagery, whereas accident-related use requires models that remain informative when vehicles are damaged, occluded, or poorly illuminated. This paper reports a four-class vehicle type classification experiment that uses transfer learning with the Visual Geometry Group 16-layer network (VGG16). A curated corpus of 1,750 images was assembled from publicly accessible web sources through a manually supervised image-retrieval process and was manually annotated before being divided into training, validation, and testing partitions (1,260 / 140 / 350). Augmentation was applied only during training. On the held-out test set of 350 images, the classifier attained 89% overall accuracy and a macro-averaged F1-score of 0.89. Class-wise F1-scores were 0.95 for bike, 0.88 for bus, and 0.87 for both car and truck. Training curves show mild overfitting on the validation split, while the separate test partition confirms balanced four-class performance. The study is limited to a VGG16 baseline; comparisons with residual and transformer architectures, together with robustness and latency tests, remain future work. Downstream effects on emergency response or smart city operations were not measured. The contribution is a transparent VGG16 baseline for four-class vehicles typing, a task-aware literature synthesis, and a clear agenda for subsequent refinement.