Accurate intraoperative tracking of keypoints is vital for trajectory planning in autonomous robotic surgery. Traditional methods for keypoint detection often rely on heatmap-based techniques, which can be computationally intensive and may not seamlessly integrate into real-time surgical workflows. In this study, we introduce a deep learning-based method for keypoint detection on soft tissue, modifying and optimizing YOLO-v8n for efficient localization of grasping and clipping points during LC using ex vivo porcine gallbladder tissues. Distinct from YOLO-v8n-pose, which focuses on learning keypoint heads, our model is trained to produce bounding boxes with their centers corresponding to the keypoints. Our approach has demonstrated an accuracy of 0.886, defined as the proportion of predicted points that fall within a defined range of the ground truth while achieving a performance of around 120 frames per second (FPS). It also outperformed previous keypoint detection models and achieved a notable reduction in computational time. These results endorse the feasibility of integration into surgical guidance systems for enhancing minimally invasive procedures in real time.