To address the difficulty of stable impact-point acquisition caused by blurred boundaries of stacked rocks, contact adhesion between adjacent targets, and complex occlusion in autonomous rock-breaking scenarios of driverless excavators, a two-stage vision-based method integrating instance segmentation and impact-point localization is proposed. The method realizes the integrated processing of rock instance perception, automatic impact-point generation, and keypoint prediction, providing a reliable visual perception basis for impact planning in autonomous rock-breaking operations. First, an improved instance segmentation model, MineBOC-YOLO, is constructed using YOLOv8-seg. Three modules are designed, namely the boundary-texture decoupling module BTDMC2f, the object-aware gated aggregation module OAGM, and the contact-relationship separation module CRSM. These modules enhance boundary-texture representation, key feature aggregation, and contact-region separation, respectively, thereby improving instance segmentation quality in densely stacked and occluded scenes. Subsequently, rock impact points are automatically calculated from instance segmentation masks, and single-keypoint annotations are generated. YOLOv8-pose is then used to predict the impact points. The experimental results show that the mask precision, mAP50, and mAP50–95 of MineBOC-YOLO reach 0.846, 0.914, and 0.683, respectively. Compared with YOLOv8n-seg, mAP50 and mAP50–95 are improved by 2.40% and 3.50%, respectively, while the number of model parameters increases only from 3.264×10? to 3.407×10?. In the impact-point localization experiment, the average errors of groups A and B are 7.07 px and 7.98 px, respectively, both lower than the 10 px threshold. The proposed method realizes the automatic transformation from instance segmentation results to impact-point localization results. While maintaining lightweight design and real-time performance, it can output rock impact positions for autonomous rock-breaking impact planning, thereby providing visual perception support for autonomous rock-breaking operations.