A multi-scale Siamese network binocular sparse point cloud generation method for visual perception in industrial IoT

In the industrial IoT environment, visual sensors, as key sensing nodes, are widely used in smart manufacturing, equipment monitoring, and digital twin scenarios, placing higher demands on the accuracy and robustness of 3D reconstruction. Binocular stereo vision is an important technical approach in the field of 3D reconstruction, and the quality of point cloud data, as a carrier of 3D spatial information, directly affects subsequent application outcomes. To address the shortcomings of traditional SURF algorithm in rotation invariance, affine transformation robustness, and illumination adaptability, this paper proposes an improved SURF feature matching method based on a multi-scale Siamese network. First, binocular camera calibration and image rectification are performed. The improved SURF algorithm is then used to detect and describe feature points in left and right images. A multi-scale Siamese network is employed to learn deep descriptors to replace traditional Haar wavelet descriptors. Experimental results show that in self-acquisition scenarios, the proposed method generates 3708 feature points in the left image, 2528 in the right image, 1354 matching points, and 237 inliers. In the Vehicle scene, our method generates 206 inliers and 108 valid 3D points, outperforming SURF (129 inliers, 58 points) and SURF+Siamese (98 inliers, 105 points). The relative depth dispersion is reduced from 30.96% (SURF) to 29.96% (ours). Furthermore, under challenging affine transformations (shear 0.2), our method produces 95 matches with 31 inliers, outperforming SIFT (42/17) and SURF (47/24), demonstrating superior robustness in industrial IoT applications.