In the fields of aviation, aerospace, automotive manufacturing, and beyond, industrial robots undertake vital tasks such as automated assembly and meticulous workpiece sorting, relying on precise recognition. Yet, in industrial environments, challenges like intricate backgrounds, workpieces with similar shapes and textures, and stacked arrangements often heighten the vulnerability to incorrect recognition. To surmount these challenges, a binocular structured light camera is first used to acquire of high-fidelity three-dimensional point cloud data. A preprocessing algorithm is then deployed to effectively eliminate background interference and mitigate noise within these complex settings. Subsequently, an innovative point cloud over-segmentation algorithm is proposed, which integrates the moving least squares method to construct local differential geometric constraints. This enables shape-preserving simplification on the raw point cloud, optimizes the supervoxel clustering process, enhances robustness against noise, and effectively mitigates point cloud adhesion. Further, a multi-feature adaptive supervoxel fusion mechanism based on concavity-convexity constraints is designed. This mechanism comprehensively integrates multi-dimensional constraints, including the concavity-convexity relationships between supervoxel clusters and geometric feature similarity, achieving high-precision instance segmentation of multi-class target in complex stacked scenarios. Building upon this foundation, a support vector machine classification architecture driven by "local-global" descriptor sequences is proposed. This architecture constructs a multi-scale cascaded feature description system that jointly characterizes the local geometric details and global morphological features of the targets. This approach effectively solves the misclassification problem caused by stacked targets under small-sample conditions. Finally, an industrial robot platform is devised for algorithm validation. Experimental findings showcase remarkable enhancements in instance segmentation accuracy and classification precision, particularly for similar weak-textured workpieces. The achieved workpiece segmentation accuracy and recognition precision surpass 95%, affirming the effectiveness and robustness of the proposed method.