A multi-stage framework for vehicle detection and classification from low-resolution surveillance images under challenging environmental conditions
TL;DR
Experimental results demonstrate that the proposed framework achieves high classification performance while maintaining an effective balance between detection accuracy and computational efficiency, indicating that the proposed method may offer a practical approach for ITS applications involving low-resolution surveillance imagery under challenging environmental conditions.
Abstract
Vehicle detection and classification play a crucial role in the effective implementation of Intelligent Transportation System (ITS) applications. Due to their superior feature representation capabilities, convolutional neural network (CNN)-based deep learning models have become strong candidates for vehicle detection and classification tasks in ITS applications. However, their computational complexity and dependence on high-resolution data acquired from advanced monitoring systems present significant challenges for deployment on resource-constrained edge devices. Therefore, to address the challenges of ITS applications, this study proposes a lightweight vehicle detection and classification framework using low-resolution surveillance images captured by standard security cameras. The proposed framework integrates transfer learning with EfficientDet1, a custom CNN, and an XGBoost classifier within a multiprocessing architecture to facilitate efficient vehicle detection and classification under varying lighting and weather conditions. To support the proposed framework, two new datasets collected under diverse environmental conditions are introduced. The first dataset consists of low-resolution vehicle images (100 × 100 pixels, 96 dpi) for classification, while the second dataset comprises annotated traffic scene images in PASCAL VOC format for vehicle detection. In addition, the publicly available DAWN dataset was utilized to further evaluate the robustness of the proposed framework under adverse weather conditions. Experimental results demonstrate that the proposed framework achieves high classification performance while maintaining an effective balance between detection accuracy and computational efficiency. The findings indicate that the proposed method may offer a practical approach for ITS applications involving low-resolution surveillance imagery under challenging environmental conditions.