Fúze 2D a 3D snímání pro rozpoznání drobných lesklých objektů

Abstract

At the present time, many developed approaches address object recognition in both 2D and 3D data. Effective object detection has become an integral part of automated operations, where machine vision systems are typically utilized. The various approaches differ in the principles of data acquisition and subsequent processing, and they also vary based on the specific purpose of the system. For objects characterized as glossy, transparent or small, the tasks of recognition, detection, and identification are significantly more complicated.

The main goal of this work is to propose a methodology for fusing 2D image data from a~camera with 3D data captured by an industrial scanner to improve the accuracy of detecting the desired objects. The specific features of the detected objects include a glossy surface and small dimensions. The work emphasizes the elimination of faults in the data caused by reflections from glossy surfaces. The fusion of 2D and 3D imaging achieves a substantial increase in the quality of the resulting detections, which in turn enhances the robustness of the detection system.

The developed methodology was applied and verified in a practical task of detecting end gauges. This task was designed in connection with the OPPIK research project OPPIK – Application – CZ.01.1.02/0.0/0.0/21_374/0026776, in which the author of this work was a co-researcher. The project focused on automating the process of removing small glossy objects from a loading cassette to the input positions of subsequent technological processes. For the purposes of this work, the project task has been generalized to not consider the predictable position of the objects of interest in the scanned area. The methodology for fusing 2D and 3D imaging is thus verified on a set of glossy objects with defined shapes and small dimensions in a undefined position.

The verification of the developed methodology is conducted by comparing the quality of detection results from 3D data with those from the fusion evaluation. The quality of detection is assessed using the FScore qualitative metric. Through a statistical test of the detection quality results, an alternative hypothesis was accepted, indicating a statistically significant increase in detection quality when using fusion evaluation of 2D and 3D data compared to using just 3D data.

The application goal of this work is to propose a robust solution for detecting small glossy objects in an industrial environment using modern technological approaches, in connection with the practical task of verifying the developed methodology. The application goal focuses on the usability of the fusion evaluation output of the input data for automated object picking applications, known as Bin-Picking, implemented in modern operations.

In the context of detecting and recognizing the picked objects, this dissertation has great significance in the potential to extend automation processes to include the manipulation of glossy objects.