A Modular Multi-Stage Approach for Classifying Low-Quality Images of Vehicles

Vehicle detection and classification are essential for effectively deploying Intelligent Transportation System (ITS) applications. While convolutional neural network (CNN)-based deep learning models deliver high performance in this domain, their computational complexity and the high-resolution data from modern surveillance cameras pose significant challenges for resource-limited edge devices. This study proposes an efficient vehicle detection and classification approach to address real-time ITS constraints using low-resolution images from standard security cameras. The method integrates transfer learning with EfficientDet1, a modified CNN, and an XGBoost classifier within a multiprocessing framework to ensure robust performance across varying lighting and weather conditions. Additionally, two new datasets are introduced: vehicle images for classification and PASCAL VOC format for traffic scene detection. The classification model is first evaluated on its dedicated dataset, followed by benchmarking detection performance against established CNN-based models, including YOLO, SSD MobileNet, SSD ResNet, and Faster R-CNN, using accuracy and computational efficiency metrics. Experimental results demonstrate that the proposed method achieves superior classification accuracy while maintaining an optimal trade-off between detection performance and computational efficiency. These findings suggest the proposed approach is a promising solution for real-time ITS applications under diverse imaging and environmental conditions.