Fujitsu Develops AI Technology to Automatically Differentiate Between Work Tasks in Video Data for Employee Training and Quality Control Purposes

TOKYO, Feb 18, 2021 – (JCN Newswire) -Fujitsu Laboratories Ltd has actually revealed the advancement of a technology that automatically spots various work components–for circumstances, “removing a part,” “tightening a screw,” or “installing a cover” – from video data including a collection of tasks from production lines and various other locations of manual labor.

Conventionally, in work environments like manufacturing facilities, initiatives to boost performance and improve quality have actually consisted of taping the work of team and determining and surpassing issues in various work components from the data. To determine the moment needed for each work component from the video data, nevertheless, it’s essential to divide video data right into different work components by hand. This commonly suggests that the man-hours needed to determine the moment for each work component is numerous times to numerous 10s of times greater than the complete size of the video.

Fujitsu has actually increased upon its “Actlyzer” technology for finding human activity and habits from video data, and created an AI version that considers the variant of each activity and the distinction in specific employees’ motions, making use of work data from a single person and the data split by each work component as training data. When this technology was used to the evaluation of procedures at Fujitsu I-Network Systems Limited’s Yamanashi Plant (Location: Minami Alps City, Yamanashi Prefecture), it was verified that work tasks can be spotted with 90% precision and that the outcomes can be made use of to boost the performance of procedures evaluation.

By leveraging this technology, Fujitsu will certainly proceed to add to the promo of work procedure renovation tasks and the handing down of details abilities for much more effective operate at numerous websites.

Background

As Japan and lots of various other innovative economic situations face decreasing birthrates, the aging of culture, and the raising movement of skill, an expanding demand exists in work that calls for human labor to document and assess important knowledge in the area and utilize it for producing handbooks and supplying onsite direction. As a component of kaizen, or work procedure renovation, tasks focused on boosting performance and quality at producing websites, several approaches of evaluation using video data (Fig1) are performed, consisting of establishing means to remove waste, contrasting deal with proficient employees, and gauging the result of time decrease via improving work procedures. In these evaluations, the work of damaging down video data of various work moves right into different work components s (Step 2 in Fig.1), consisting of dividing the begin and end of a component, commonly takes greater than one hr for a 20 min video, for instance. This difficult editing and enhancing procedure has actually offered obstacles to initiatives to increase work performance and the transmission of abilities.

Issues

In current years, it has actually come to be feasible to find the activity of the human skeletal system from video photos and to identify usual tasks done by several individuals by training the system making use of deep discovering innovations. In order to attain this, nevertheless, it’s essential to consider the refined variants in motions that take place between people also when they carry out the exact same tasks, in addition to the distinctions in motions amongst various individuals and settings. This need to be done while furthermore comparing comparable motions and preparing big quantities of training data.

Newly Developed Technology

Fujitsu has actually efficiently created a technology to automatically find work components also for various video data of the exact same work, thinking about variants in activity and various other comparable motions, by finding out each work component making use of data acquired by separating a solitary video for each component as training data.

First, in order to train the formula for various work components, three-dimensional skeletal system acknowledgment technology consisted of in Fujitsu’s habits evaluation technology “Actlyzer”( 1) is made use of to approximate an individual’s pose in three-dimensional room from the picture that ends up being training data. Then, adjustments in the pose of the top body are gotten as attribute worths at periods of numerous hundred nanoseconds to numerous secs, and comparable attribute worths are organized right into numerous 10s of system motions (m1-mn in Fig.2). In this situation, if the features of the modification in pose are comparable, it is considered the exact same activity, so it is feasible to take care of the distinction in the method it shows up depending upon the functioning setting and the setting of the video camera.

Next, the department setting (timestamp) of the work component area equivalent to the training data video is made use of to connect the very discreet activity with the work component. At this time around, by approximating the modification pattern of the mix of very discreet system motions that make up the work component and producing it as an AI version, it is feasible to soak up the distinction in the activity of people and the variants in the great activity that takes place whenever also when the exact same work component is done. In enhancement, the AI version automatically acknowledges work components from the series of system motions. This makes it feasible to discover the work component that ideal considers the total order of work, also if the exact same work component exists greater than as soon as, such as “tighten a screw”.

Outcomes

At the Yamanashi Plant of Fujitsu I-Network Systems, which produces network tools items, Fujitsu used the established technology to assess the complying with 3 work procedures (Fig3): components establishing, setting up, and aesthetic assessment, and carried out an examination.

In each procedure, by just training with the department data of each work component of a single person’s video, the work component is automatically spotted also in an additional video of the exact same deal with 90% or even more precision. As an outcome, the cycle of kaizen tasks can be duplicated much more often, assisting to boost work performance and increase the transfer of abilities.

Future Plans

In enhancement to the production sector, Fujitsu will certainly proceed to confirm this technology in numerous usage situation situations, consisting of logistics, farming, and healthcare, with the goal of placing it right into functional usage within financial 2021.

( 1) behavior evaluation technology “Actlyzer”

Deep finding out technology that can identify a range of refined and complicated human motions without depending on big quantities of training data. (Press Release: 2019/11/25)

About Fujitsu

Fujitsu is the leading Japanese details and interaction technology (ICT) firm supplying a complete variety of technology items, remedies and solutions. Approximately 130,000 Fujitsu individuals sustain consumers in greater than 100 nations. We utilize our experience and the power of ICT to form the future of culture with our consumers. Fujitsu Limited (TSE:6702) reported combined profits of 3.9 trillion yen (US$ 35 billion) for the finished March 31, 2020. For even more details, please see www.fujitsu.com.

About Fujitsu Laboratories

Founded in 1968 as an entirely had subsidiary of Fujitsu Limited,Fujitsu Laboratories Ltd is just one of the leading proving ground in the globe. With an international network of labs in Japan, China, the United States and Europe, the company carries out a vast array of standard and used study in the locations of Next- generation Services, Computer Servers, Networks,Electronic Devices and Advanced Materials For even more details, please see: http://www.fujitsu.com/jp/group/labs/en/.

Source: Fujitsu Ltd

Sectors: Cloud & Enterprise