A Lightweight Real-Time Vehicle Make and Model Recognition Method

Recently, Vehicle Make and Model Recognition (VMMR) has attracted widespread attention in the areas of intelligent transportation and computer vision. However, large deep convolutional neural networks (CNNs) often prove impractical on edge devices with limited computing power due to large model size and excessive inference time. Inspired by facial recognition, this paper introduces a lightweight “vehicle face recognition (VFR)” approach designed for parking lot scenarios, aiming to achieve real-time and accurate recognition of vehicles' make and model on mobile devices. We focus on the vehicle front view and select the Frontal-103 dataset for experimentation, with all images annotated and preprocessed. Additionally, this paper introduces an improved MobileNetV2 model that achieves a recognition accuracy of 94.76% on the test dataset. Finally, the MobileNetv2-Improved is deployed on the low-power embedded device RK3568 for testing. The results indicate that the improved model can achieve superior accuracy while maintaining remarkable real-time performance.