【Purpose】In order to solve the problems of target smoke misdetection and low detection accuracy in complex scenes.【Method】Improve and optimize the YOLOv7 algorithm which was on the current best performing object detector, replaces the PAFPN structure in the Neck part of the original model with the asymptotic feature pyramid structure AFPN and uses ECIoU as the objective regression loss function, and verifies it on the selfconstructed dataset SM-datase and Pycharm platform.【Result】Experimental results show that the accuracy of the improved algorithm was increased by 1.3% to 68.6% compared with the original YOLOv7 model, the average accuracy mAP is increased by 1.8% to 64.6% compared with the original YOLOv7 model, and the computational complexity of the improved algorithm is only 82.5 GFLOPs, which was 27.4% lower than that of the original YOLOv7 model.【Conclusion】Based on the improved YOLOv7 algorithm, the algorithm proposed in this paper can not only reduce the computational complexity of the network but also improve the detection accuracy in complex scenes, which provides a new idea for the follow-up research of smoke detection in complex scenes.