Search

Since traditional pose estimation methods with the features of points, lines and so on might not be applied directly to round-shape workpieces, a new pose estimation method for round-shape workpieces genetic algorithm based was proposed. Compared with previous studies, this method needs no auxiliary information, such as points, lines, concentric circles and so on. Firstly, transformation model of perspective projection of round-shape workpiece was created, and the round-shape workpiece was characterized by analytic equation. Secondly, via detecting the contour and extracting its feature of workpiece, a feature error function was established with respect to the pose angles, which was a multi-objective nonlinear function. Finally, the error function was solved by an improved genetic algorithm and the pose estimation of round-shape workpiece was achieved. The result of related experiments showed that the method had high accuracy, and to some extent inhibited the effects of noise.

In the picking robot binocular vision systems research, the camera calibration is often an indispensable step and these basements to locate the target of the object and rebuild the three-dimensional construction based on the robot stereo vision for the follow-up study. So, searching for a high accuracy and simple camera calibration algorithm is of great significance and necessary. However, For most of these camera calibration algorithms, it is necessary to establish a reference object, namely the target, in front of the camera at present, but posing the target is very not convenient or almost impossible in some cases. Therefore, a picking robot online calibration algorithm based on the vision scene was proposed by studying the work environment characteristics of the picking robot binocular vision system and the invariant projective geometry. The experimental results showed that this algorithm’s calibration accuracy and precision good meets to the requirement of the robot binocular vision system camera calibration in the complex environment.

Aiming at the problem of low recognition rate and easy affected by environment during the process of robot target recognition in complex environments, the target recognition method combining support vector machine (SVM) with D-S evidence theory was proposed. Taking citrus recognition as an example, SVM was used by the method to local classification according to citrus color and geometry feature information respectively, and the results of SVM were transformed to probability outputs through Platt model, and treated them as the basic probability assignment (BPA) of D-S evidence theory to reason and fuse local recognition results, and then realized the combination of SVM and D-S evidence theory in citrus recognition, finally improved the recognition rate. The experimental results showed that: the recognition rate of the method combining SVM with D-S evidence theory and integrating color features and geometry features was higher than SVM method with only color or geometry features.