Giving artificial intelligence to industrial vision systems – a cooperation with Caterpillar
In this presentation we will show that visual perception for industrial processes could benefit greatly from the current interest in applying autonomous technologies to automated driving systems. The artificial intelligence (machine learning) that is the basis for the visual perception of these automotive systems could for instance, improve visual quality inspection. It could enhance or replace a human visual inspection, and as such, optimize tedious, time-consuming and error prone tasks. For instance, Caterpillar automated ground-truthing and data labelling, which dramatically limited the need for human supervision and significantly reduced development time.
However, developing and integrating such a solution in an industrial environment, poses multiple challenges to engineers and domain experts. These challenges can be grouped in three categories. First, especially for deep learning, you require massive amounts of image data for training and testing. Next to that it requires expertise to fine-tune these complex models. Finally, you need to leverage high-performance computers and maybe deploy the technology into embedded devices like GPUs.
These challenges were also true for Caterpillar, the world leading manufacturer of construction and mining equipment. The environments where these multimillion dollar machines operate can be harsh and dangerous, especially for the human operators. Because of this, Caterpillar is working on Autonomy and Operator Assistance. In this presentation we show that they considered big data and machine/deep learning processing, but were spending too much time on ground truth labelling and managing training and testing data. To address this, they developed an infrastructure to integrate the entire workflow.Slides not available - contact speaker