Multi-Model Evaluation of Computer Vision Techniques for Fine-Grained Vehicle Recognition

Abstract

Vehicle make and model recognition is a crucial component in applications such as intelligent transportation systems, law enforcement, and autonomous vehicles.This paper presents a comparative analysis of Machine Learning approaches — KNN, SVM and Decision Tree Classifier which are the first appraoch and three deep learning models—Convolutional Neural Network (CNN), YOLOv8, and Faster R-CNN— are the second approach for vehicle recognition tasks. The models were evaluated on a two datasets comprising 197 vehicle classes (dataset1) under various conditions and the other with was scrapped manually with 17 classes (dataset2), focusing on metrics such as validation accuracy, inference speed, robustness to occlusion, and computational efficiency. YOLOv8 emerged as the best-performing model in both the datasets, achieving a mean Average Precision (mAP) of 95% with dataset1 and 30% with dataset2, with an inference speed of just 10 ms per image, making it highly suitable for real-time applications. Faster R-CNN demonstrated exceptional precision and robustness in handling complex scenarios but was constrained by slower inference speeds with an accuracy of 74% with dataset1 and 65% with dataset2. In CNN, which is  computationally efficient, suffered from significant overfitting, with a testing accuracy of only 10% with dataset2. Whereas with dataset1 they gave an 78% of testing accuracy and 90% of training accuracy. The findings of this study emphasize the strengths and limitations of each model, providing insights into their applicability across various real-world scenarios.