Integrating CNN and RANSAC for improved object recognition in industrial robotics

Abstract

This research introduces a robotic grasping system that merges ORB (Oriented FAST and Rotated BRIEF) feature detection, VGG19 convolutional neural networks, and RANSAC (Random Sample Consensus) geometric verification to achieve high-precision object manipulation in unstructured environments. The framework synergizes ORB's efficient, rotation-invariant keypoints with deep semantic features extracted from intermediate layers of VGG19 enabling robust object recognition under occlusions and lighting variations. ORB detects scale-agnostic keypoints and generates binary descriptors, while VGG19’s hierarchical features provide contextual understanding of object geometry. These complementary features are fused into compact descriptors, combining ORB's 256-bit binary patterns with aggregated VGG19 layer outputs to balance accuracy and computational efficiency. RANSAC is then employed to eliminate mismatched features and estimate precise spatial alignments through iterative homography calculations, ensuring reliable mapping between detected objects and the robot's workspace. Experimental validation on industrial dataset trials demonstrates a 99 % grasp success rate, highlighting the system's ability to address challenges in dynamic, cluttered settings. By bridging deep learning's perceptual capabilities with geometric verification, this work advances autonomous robotic systems, offering a scalable solution for industrial automation that prioritizes precision and adaptability.