The widespread development of autonomous driving has posed a significant challenge in traffic light detection. Existing detection technologies often fail to effectively balance the real-time performance and accuracy of object detection. This paper proposes a new object detection model, MRT-DETR, based on the DETR series, which effectively achieves a balance between real-time performance and accuracy in traffic light detection. MRT-DETR employs depthwise separable convolutions and optimize the receptive field in certain layers of the backbone network. Also, MRT-DETR replaces the traditional ReLU activation function with the HardSwish activation function. To demonstrate the effectiveness of the MRT-DETR model, this paper conducts experiments on a traffic light detection dataset, comparing it with common object detection models, including CNN-based models and the DETR series, and analyzes the results. The experimental results show that the MRT-DETR model achieves an 80.61% improvement in the \(\text {AP}^{val}\) metric and a 68.42% improvement in the \(\text {AP}^{val}_{50}\) metric compared to the YOLOv10l model. Compared to the RT-DETR model, MRT-DETR improves the \(\text {AP}^{val}\) metric by 3.22% and the \(\text {AP}^{val}_{50}\) metric by 16.93%. In terms of inference time, the MRT-DETR model reduces inference time by 7.8% compared to the RT-DETR model. The experiments show that the MRT-DETR model achieves favorable results in balancing the accuracy and real-time performance of object detection.