In view of the complex urban environment, the rescue after the earthquake has the problems of not timely discovery and low recognition rate. A small target human detection algorithm model SE-YOLOv7 based on YOLOv7 is proposed to solve this problem. Based on YOLOv7, a human body recognition data set is made for complex urban disaster areas, and the residual network module SE attention mechanism is added to the framework to improve the contribution of positive sample features to better train the model weights. Combined with the characteristics that the affected human body imaging is different from the normal situation, the distribution density of the length and width of the prior box is improved, and the model is deployed on the embedded platform of the post-disaster eight-legged rescue robot[3]to carry out the human target detection experiment in the post-disaster environment. The accuracy of human limb recognition is2.3%map higher than that of voc2012 data set, the recall rate is increased by 2.7%, and the speed is 67.6 FPS, which meets the real-time and accuracy requirements of target detection of rescue robots under disaster conditions.