Improved RT-DETR Based on MobileNetV4 for Vehicle Detection

In this study, we present an enhanced version of the RT-DETR model integrated with MobileNetV4, designated as RT-DETR-MobileNetV4-Small/Medium, tailored for efficient vehicle detection. This model enhances computational efficiency while retaining high detection accuracy, making it ideal for deployment in resource-constrained environments. Key innovations include a streamlined backbone architecture using MobileNetV4 that significantly reduces parameter count and computational demands. Extensive evaluations on a COCO-format vehicle dataset demonstrate the model's superior performance. Specifically, our RT-DETR-MobileNetV4-Small achieves accuracy comparable to, and in some cases superior to, other RT-DETR models while utilizing only 11M parameters and 38 GFLOPs. This represents a reduction of 65% in parameters and 75% in computational complexity compared to RT-DETR-L, and a 45% reduction in parameters and 54% reduction in GFLOPs compared to RT-DETR R18. These results highlight the potential of RT-DETR-MobileNetV4 in applications requiring real-time processing on mobile and edge devices, offering a promising solution for advanced object detection tasks in dynamic environments.