Computer vision software for robots and other machines

Rtrilo is computer vision software that combines deep learning technologies with various types of image processing technologies. Capable of identifying specific objects and areas from both videos and stills, it enables high-level control of robots. The software provides robots and various other machines such as autonomous carts with the functions of “sight.”

Overview of Rtrilo

Providing robots and other machines with “sight”

“Computer vision” refers to technology that identifies objects in videos and still images, thereby providing computers with visual functions. Rtrilo combines image processing and deep learning technologies to detect and identify specific objects—tasks that had proven difficult for existing image processing software. In addition, the software leverages unique image processing technologies to provide situational awareness that can be applied to a variety of different scenarios.

Usage scenarios

Rtrilo can be equipped to robots and various other machines to detect specific objects and obstacles, identify no-entry areas, and grasp the surrounding environment. In this way, the software provides machines with the functions of “sight.” It can be used in the following scenarios.

Detecting travel routes and obstacles

Rtrilo detects no-entry areas and travel routes—which distance sensors and other devices find it hard to differentiate between— in real time, enabling robots to travel flexibly and autonomously.

Recognition of specific objects

Rtrilo is capable of remotely recognizing objects with specified colors and shapes, and enabling robot movements to be controlled accordingly. In the image, Rtrilo is tracking the human holding the blue case.

Identifying how to grip different objects

By incorporating deep learning models that have been learned via unique methods, Rtrilo can identify how to grip different tableware, and differentiate between agricultural grains and weeds.

Reading seven-segment displays

Rtrilo is able to recognize a variety of information, such as warning lamps and seven-segment displays, in camera images, and send real-time alerts to managers or other destinations. This is useful for existing devices that, due to various constraints, cannot easily be connected to IoT.

Rtrilo Learning Box: deep learning model construction tool

 


The construction of deep learning models for specific object detection and image classification requires a variety of processes, including complex pre-processing, defining of deep learning models, adjustment of learning parameters, and performance evaluation. Rtrilo Learning Box is equipped with the functions required to construct such models and so enables customers to easily construct deep learning models themselves.

 

■ Types of deep learning models that can be constructed

  • Image classification: the ability to differentiate between what is being shown in various images
  • Object recognition: the ability to detect different objects in an image, and differentiate between them

 

■ Pre-processing

  • Annotations (creation of teaching data)
  • Division of data (training, verifying, testing)
  • Image amplification (noise, cropping, color changes, blurring, contrast changes, rotation, geometric changes, etc.)

 

■ Learning

  • After various parameters have been specified, including the model to be applied, imbalance corrections, learning cycles, feature scaling, etc., learning programs can be executed with a single command line.
  • Loss functions, optimization algorithms, learning rates, and other learning methods can be freely adjusted, and unique network models can be constructed.

 

■ Evaluation

  • Output learning curves, various indices (precision and recall ratios, f-scores), and confusion matrices
  • Output erroneously judged images

Documents