To address the problems of light interference, target occlusion, wide range of target scale variation and uneven quality of training images during fall, a fall recognition algorithm based on an improved YOLOv7 model is presented in this paper. By incorporating Coordinate Attention mechanism in Backbone, the system pays attention to the location information of the fall target. And the InceptionNeXt module is employed, which aims to maximise the feature information extraction, to enhance the accuracy of target detection at different scales and improve the performance when there are light interference and occlusion situations in the environment. In addition, the loss is replaced with Wise-IoU, which is applied to address the negative impact of low-quality images on the model performance and obtain more accurate findings. The findings of the experiments indicate that the improved models mAP0.5 and mAP0.5:0.95 in this paper improve by 1.7% and 6.4%, respectively, compared to the original YOLOv7 model. Therefore, the model presented in this paper has an advanced performance in the fall detection task.