Accurate weld-region identification is essential for weld quality inspection and automated grinding. However, weld point clouds are highly irregular and lack explicit topological structure, which makes accurate recognition challenging. To address this issue, this study formulates weld point-cloud recognition as a binary point-wise classification task. Each point is classified as either weld bead or base metal. A systematic classification framework is established by combining neighborhood-based geometric feature extraction, baseline model comparison, and metaheuristic hyperparameter optimization. Three morphology-specific weld subsets, including straight-line, curved-line, and S-shaped welds, are used for evaluation. The classification performance of Random Forest (RF), Extreme Gradient Boosting (XGBoost), and Light Gradient Boosting Machine (LightGBM) is first compared under different neighborhood scales. Overall Accuracy (OA), Precision, Recall, and F1-Score are used as evaluation metrics. The results show that LightGBM achieves the best baseline performance at a neighborhood radius of 1.5 mm. To further improve classification performance, LightGBM hyperparameters are optimized using metaheuristic algorithms. The compared optimizers include the Artificial Lemming Algorithm (ALA), Alpha Evolution Algorithm (AE), and Starfish Optimization Algorithm (SFOA). Repeated-run results demonstrate that AE-LightGBM achieves the most favorable overall performance under the unified evaluation protocol. Statistical significance analysis and convergence analysis further support the effectiveness of AE among the compared optimizers. In addition, SHapley Additive exPlanations (SHAP) is employed to analyze feature contributions and improve the interpretability of the optimized model. The proposed method provides an effective technical pathway for robot-based weld recognition and grinding tasks using 3D vision and supervised machine learning.

Accurate identification of weld regions is a critical step in weld quality inspection and automated grinding operations. However, weld point cloud data is typically characterized by high irregularity and lack of topological structure, making its efficient processing and accurate recognition a persistent research challenge. In recent years, advanced machine learning methods have shown significant potential for weld identification, yet research on 3D vision-based recognition of three-dimensional weld point clouds remains in its early stages. This study focuses on three typical weld seam trajectories—linear, curved, and S-shaped—conducting classification and recognition experiments based on point cloud data. Using Overall Accuracy, Precision, Recall, and F1-score as evaluation metrics, a systematic comparison was performed on the classification performance of three models—Random Forest (RF), Extreme Gradient Boosting (XGBoost), and Light Gradient Boosting Machine (LightGBM)—across different neighborhood scales. Experimental results indicate that LightGBM achieved optimal classification performance within a 1.5mm neighborhood range. To further enhance the classification accuracy, this study introduced the Artificial Lemming Algorithm (LAL), Alpha Evolution Algorithm (AE), and Starfish Optimization Algorithm (SFOA) to improve the LightGBM model. The optimized LAL-LightGBM, AE-LightGBM, and SFOA-LightGBM models attained accuracy rates on linear, curved, and S-shaped test sets of (98.02%, 96.07%, 93.87%), (98.83%, 96.36%, 94.18%), and (98.00%, 96.12%, 93.91%), respectively. Comprehensive results demonstrate that the AE-LightGBM model performed best in overall evaluation, verifying the effectiveness of intelligent optimization algorithms in improving classification model performance. Furthermore, this study employed the SHAP interpretability framework to analyze the feature contributions of the AE-LightGBM model, effectively revealing its decision-making mechanism and enhancing model interpretability. This research provides a feasible technical pathway for robot-based weld grinding recognition tasks utilizing 3D vision and supervised machine learning.

In deep-learning-based multi-object tracking frameworks, deep neural networks are used for object detection, appearance feature extraction, and data association, while filtering models perform target-state prediction and observation update for continuous trajectory estimation. Under occlusion, overlap, and false detections, matched detection boxes may exhibit position shifts and scale perturbations, causing inconsistency between current observations and prior predictions and leading to identity switches, trajectory fragmentation, and slow tracking recovery. To address these problems, this paper proposes an improved Gaussian Strong-Tracking Adaptive Kalman Filter (GSTAKF). The proposed GSTAKF jointly uses detection confidence, filtering residuals, and the strong-tracking factor to adaptively regulate the observation-noise covariance and state-update strength, thereby improving the response to low-reliability observations and abrupt motion variations. After appropriate framework-specific adaptation of the state representation, covariance scales, and filtering parameters, GSTAKF can be integrated into deep-learning-based SORT-type multi-object tracking algorithms as a modular plug-and-play motion-estimation component. Experiments on MOT17 show that under position perturbations of 0–25 pixels and scale perturbations of 0–40%, GSTAKF improves HOTA by 18.725, 17.096, 7.875, and 1.805 points when integrated into DeepSORT, StrongSORT, BoT-SORT, and HybridSORT, respectively. On MOT20, under the same perturbation ranges, GSTAKF improves HOTA from 49.022 with Standard KF to 67.328, corresponding to a gain of 18.306 points. On DanceTrack, without artificial noise, GSTAKF improves HOTA from 43.369 to 47.173, corresponding to a gain of 3.804 points over Standard KF.