To mitigate missed detections and high false detection rates caused by brightness attenuation, noise interference, and blurred target boundaries in low-illumination images, this paper introduces the MST-YOLO algorithm featuring temporal multi-scale feature reconstruction and dynamic task alignment. This method innovatively builds a multi-scale sequence feature fusion network (MesNet) that collaboratively optimizes shallow high-resolution edge features and deep robust semantic features through a cross-layer enhancement mechanism. At the same time, an NMS-free detection head is designed to eliminate the dependence on traditional non-maximum suppression and build an end-to-end reasoning process. Experiments on the DarkFace low-light dataset demonstrate that MST-YOLO achieves an of 59.5%, an 11.7% improvement over the YOLOv8 baseline, validating its dual advantages in enhancing feature representation and optimizing boundary positioning accuracy for target detection in complex lighting conditions.