Multi-state robot posture detection method based on joint optimization of target-key point detection

Introduction Industrial robots are the core equipment in intelligent manufacturing. The accuracy of their posture detection directly determines their high-precision operational capabilities. It is necessary to address the problems existing in the current visual detection methods, such as focusing only on the state of a single robot, difficulty in balancing real-time performance and accuracy, and the unbalanced optimization of multi-joint angle errors. Methods A multi-state robot posture detection method based on joint optimization of targets and key points has been proposed. This method employs an improved cross-stage local network as the backbone network, simultaneously outputting the features of target detection and key point detection, and achieving the collaborative optimization of dual tasks through weighted joint loss. Additionally, a multi-state adaptive module is designed, which determines the state based on the motion vectors of adjacent frames, compensates for motion blur using directionally fast rotation binary robust independent basic feature matching, and corrects the perspective deviation using posture correction parameters. Results The results show that when the intersection and union threshold of the introduced model ranges from 0.5 to 0.95, the average target detection accuracy is 94.73%; when the key point detection threshold is 0.2, the key point detection accuracy rate is 98.7%; the average absolute errors of joint angles in the fixed, moving and unknown states are 7.2°, 8.1° and 9.5° respectively, and the inference speed is only 43 frames per second. Discussion This method can enhance the detection performance in multi-state scenarios and provide technical support for the complex operations of industrial robots.