CSDet-3D:A novel 3D object detection method for robot identification in welding scenarios

In the process of intelligent manufacturing transformation driven by Industry 4.0, digital factories have gradually become the core carrier. As a typical application scenario of intelligent manufacturing, automotive welding assembly faces a technical bottleneck in constructing digital factories: the high-precision spatial positioning of process equipment. Welding scenarios are characterized by high complexity, a large number of process equipment, and dense layouts, making it difficult to obtain precise positioning information of key equipment (such as manipulators) through conventional methods. To address this, this paper proposes a 3D point cloud detection network for welding scenarios to achieve precise positioning of process equipment. The algorithm innovatively introduces a color feature extraction descriptor to address the weakened generalization ability caused by painting differences among process equipment of different models. Meanwhile, a shape-aware feature-based attention mechanism module is introduced to enhance the capture of key structural features of process equipment.We also collect point cloud data from a real welding production line of an automotive brand to establish a dataset for training and comprehensive verification. Simultaneously, further experiments are conducted on the public dataset S3DIS to validate the effectiveness of the method. Our method achieved mean average precision (mAP) scores of 89.5% and 53.1% on the welding point cloud dataset. Compared with the baseline model, the mean average precision (mAP) under different thresholds increased by 5.3% and 3% percentage points, respectively., providing a reliable perceptual foundation for the virtual-real mapping of intelligent factories.