On object recognition for industrial augmented reality
Título de la revista
ISSN de la revista
Título del volumen
Some reasons are market pressure, an increase of functionality, and adaptability to an already complex environment, among others. Therefore, workers face fast-changing and challenging tasks along with all the product lifecycle that reach the human cognitive limits. Although nowadays some operations are automated, many of them still need to be carried out by humans because of their complexity. In addition to management strategies and design for X, Industrial Augmented Reality (IAR) has proven to potentially benefit activities such as maintenance, assembly, manufacturing, and repair, among others. It is also supposed to upgrade the manufacturing processes by improving it, simplifying decision-making activities, reducing time and user movements, diminishing errors, and decreasing mental and physical effort. Nevertheless, IAR has not succeeded in breaking out of the laboratories and establishing itself as a strong solution in the industry, mainly because technical and interaction components are far from ideal. Its advance is limited by its enabling technologies. One of its biggest challenges are the methods for understanding the surroundings considering the different domain variables that affect IAR implementations. Thus, inspired by some systematical methodologies proposing that, for any problemsolving activity, it is required to define the characteristics that constrain the problem and the needs to be satisfied, a general frame of IAR was proposed through the identification of Domain Variables (DV), that are relevant characteristics of the industrial process in the previous Augmented Reality (AR) applications. These DV regard the user, parts, environment, and task that have an impact on the technical implementation and user performance and perception (Chapter 2). Subsequently, a detailed analysis of the influence of the DV on technical implementations related to the processes intended to understand the surroundings was performed. The results of this analysis suggest that the DV influence the technical process in two ways. The first one is that they define the boundaries in the characteristics of the technology, and the second one is that they cause some issues in the process of understanding the surroundings (Chapter 3). Further, an automatic method for creating synthetic datasets using solely the 3D model of the parts was proposed. It is hypothesized that the proposed variables are the main source of visual variations of an object in this context. Thus, the proposed method is derived from physically recreated light-matter interactions of this relevant variables. This method is aimed to create fully labeled datasets for training and testing surrounding understanding algorithms (Chapter 4). Finally, the proposed method is evaluated in a study case of object classification of two cases: a particular industrial case, and a general classification problem (using classes of ImageNet). Results suggest that fine-tuning models with the proposed method reach comparable performance (no statistical difference) than models trained with photos. These results validate the proposed method as a viable alternative for training surrounding understanding algorithms applied to industrial cases (Chapter 5).