Early fires on battery-powered ships manifest as smoke, posing higher requirements for detection accuracy, latency, and real-time detection of video streams. A real-time smoke detection model based on an improved DETR is proposed. Smoke simulation experiments are conducted in maritime engine room scenes to construct smoke datasets. Subsequently, by using a lightweight convolutional structure and improving parameter transmission, the model's smoke feature extraction capability is optimized. In the encoder part, a cross-scale feature fusion structure is constructed to achieve the self-attention mechanism model for detecting streaming videos. During the testing phase with smoke videos, a detection speed of 29 FPS at a resolution of 1080p is achieved by our proposed method, and detection accuracy metrics AP0.5:0.95 and AP0.5 reaching 80.5% and 96.9%. The experimental results demonstrate that excellent performance in real-time smoke detection for battery-powered ships is exhibited by this method.