Robot Vision

2019+

The Development of a New MOP (Mathematical
Optimization) Algorithm for Box Recognition

Recognition of boxes that are all in one shape and on a single layer at high speed without a need for pattern registration!

2016-2018

Improved Learning-Based Recognition through the Combined Use of 3D and AI (DeepLearning)

By implementing AI, it is now possible to recognize almost all boxes in a mixed state!

2015–2016

Improved Recognition Accuracy Through the Combined Use of Point Cloud Information of 3D Camera and Color Image

Using a 3D camera makes it possible to measure the shape of the top surface of boxes more accurately and at high speed, and using this in combination with a 2D color camera as well as by using contrast successfully resulted in improved box division accuracy.In addition to standard depalletization software, we implemented new robot calibration software for multiple 3D cameras and various robot manufacturers as well as a hand interference check function.

2014–2016

Color Image Processing Added to Point Cloud Processing for Bag Recognition

Unlike boxes, the boundaries (edges) of individual bags have gentle peaks and valleys. We achieved the commercialization of bag recognition by creating a mesh on a very small plane, setting the difference between normal vectors as the boundaries, and recognizing each bag in combination with segmentation and contrast.

2013–2015

Improved Accuracy with Contrast Imaging

As a method of recognizing the condition where there are multiple boxes on the same plane, there is almost no difference in height, and there is almost no gap between boxes, we adopted a method of dividing the boxes on the same plane by utilizing the difference in contrast.

2010–2013

Combining the Use of a Grid Laser and a Stereo Camera

We delivered the world’s first robot vision for box recognition that supports random workpieces!!

Imagine a situation in which there is an infinite variety of boxes that are loaded randomly on pallets or cage carts, and you have no prior information on the types of boxes that are there. “How would you conduct image recognition?”
This is the challenge our company began to take on in 2010. If you know the type of boxes being used and have information about the size and print pattern design of the boxes in advance, it can be relatively easy to recognize the individual box by using something like standard two-dimensional image processing template matching. However, in a situation where you have no prior information and you don’t know what has been loaded, how you accurately recognize individual boxes is key. In 2010, the performance of the device known as a 3D camera was inferior to that of today, but through trial and error, our company developed a method for conducting 3D measurement by combining a grid laser with a stereo camera, and introduced this method to a major distribution center.