DEEP LEARNING-BASED VEHICLE MAKE AND MODEL RECOGNITION: DATASET CHALLENGES AND INFLUENCING FACTORS IN URBAN TRAFFIC

Abstract

This paper presents a deep learning-based approach to vehicle make and model recognition (VMMR). Using ResNet50, AlexNet, VGG19, MobileNet, EfficientNet, VIT b16, ConvNeXt, SimCLR, CLIP we address the challenges of annotation of data sets, class imbalance, and quality of real-world cameras. A data set of 283,899 vehicle images was extracted and tested, achieving validation accuracy of 0.07 to 0.32. We analyze how vehicle size, acceleration, and turning radius affect detection performance and propose directions for optimizing smart city traffic systems.