Abstract: In order to help drivers on complex and variable mountain roads timely discover falling rocks in front of them, the "clustering method" was used to mark falling rocks in an innovative way, and the global maximum average pooling was embedded into the SE attention mechanism to complete the design of MA- SE module. The MA- SE module was extended to ELAN module in YOLOv7 network by means of module group, and filled into SPPCSPC module based on location advantage. Meanwhile, the IOU loss function was optimized in parallel, and a new YOLOv7 network model based on embedded maximum average pooling SE module was proposed. On the self- made rockfall data set, the improved YOLOv7 algorithm increases mAP@0.5 by 1.6% with almost no decrease in FPS. At present, there are few studies on the detection of rockfall on mountain roads, which has certain research value.