Real-Time Multi-Target Laser-Stripe Keypoint Estimation for Robotic Adhesive Dispensing Using a Two-Stage Lightweight Network

Reliable extraction of local trajectory geometry from line-laser images is essential for robotic adhesive-dispensing perception, particularly in multi-target scenes affected by reflection, stripe discontinuity, and weak responses. This study presents a two-stage framework that separates full-image target localization from ROI-level keypoint estimation. YOLO26n first detects local trajectory regions, after which an orientation-aware lightweight laser keypoint network (OLKNet) predicts one center control point and two directional reference points within each ROI. OLKNet combines a FastViT-T8 backbone, a lightweight top-down feature pyramid, center-guided rotated anisotropic heatmaps, and channel-dependent DARK decoding. A dataset comprising 4460 images and 5656 target instances was constructed. On the test subset, YOLO26n achieved a precision of 0.9878, a recall of 1.0000, and an mAP${}_{50:95}$ of 0.8737. On 567 test instances matched to predicted ROIs, OLKNet achieved an AP of 0.9120, a center-point mean Euclidean error of 3.406 pixels, a center-point root mean square error of 4.316 pixels, and an angular error of 2.819${}^{\circ }$. Under a five-frame scheduling strategy, the GPU vision-software pipeline achieved 73.93 FPS for single-target images and 47.17 FPS for multi-target images. Here, real-time performance refers to the vision-software pipeline rather than robotic closed-loop dispensing.